diff --git a/Makefile b/Makefile
index cfb1e7e..9898e4e 100644
--- a/Makefile
+++ b/Makefile
@@ -69,12 +69,10 @@ BUILDPYS= $(SAGES:test/%.sage=$(BUILD_PY)/%.py)
 .PHONY: clean all test test_ct bench todo doc lib bat sage sagetest gen_headers
 .PRECIOUS: $(BUILD_ASM)/%.s $(BUILD_C)/*/%.c $(BUILD_H)/*/%.h $(BUILD_IBIN)/%
 
-GEN_HEADERS=\
-	$(BUILD_INC)/decaf/decaf_255.h \
-	$(BUILD_INC)/decaf/decaf_448.h \
-	$(BUILD_INC)/decaf/decaf_255.hxx \
-	$(BUILD_INC)/decaf/decaf_448.hxx \
-	$( src/public_include/decaf/* : src/public_include = $(BUILD_INC) )
+HEADER_SRCS= $(shell find src/public_include -name "*.h*")
+GEN_HEADERS_0= $(HEADER_SRCS:src/public_include/%=$(BUILD_INC)/%)
+GEN_HEADERS_1= $(GEN_HEADERS_0:%.tmpl.h=%.h)
+GEN_HEADERS= $(GEN_HEADERS_1:%.tmpl.hxx=%.hxx)
 HEADERS= Makefile $(shell find src test -name "*.h") $(BUILD_OBJ)/timestamp $(GEN_HEADERS)
 
 # components needed by the lib
@@ -90,7 +88,6 @@ scan: clean
 		 -enable-checker osx -enable-checker security -enable-checker unix \
 		make all
 
-
 # Internal test programs, which are not part of the final build/bin directory.
 $(BUILD_IBIN)/test: $(BUILD_OBJ)/test_decaf.o lib
 ifeq ($(UNAME),Darwin)
@@ -125,10 +122,15 @@ $(BUILD_OBJ)/%.o: $(BUILD_ASM)/%.s
 	$(ASM) $(ASFLAGS) -c -o $@ $<
 
 gen_headers: $(GEN_HEADERS)
+
+$(BUILD_INC)/%: src/public_include/% $(BUILD_OBJ)/timestamp
+	cp -f $< $@
+	
+$(BUILD_INC)/%.h: src/public_include/%.tmpl.h src/gen_headers/*
+	python -B src/gen_headers/template.py --per=global --guard=$(@:$(BUILD_INC)/%=%) -o $@ $<
 	
-$(GEN_HEADERS): src/gen_headers/*.py src/public_include/decaf/*
-	python -B src/gen_headers/main.py --hpre=$(BUILD_INC) --ihpre=$(BUILD_H) --cpre=$(BUILD_C)
-	cp src/public_include/decaf/* $(BUILD_INC)/decaf/
+$(BUILD_INC)/%.hxx: src/public_include/%.tmpl.hxx src/gen_headers/*
+	python -B src/gen_headers/template.py --per=global --guard=$(@:$(BUILD_INC)/%=%) -o $@ $<
 
 ################################################################
 # Per-field code: call with field, arch
@@ -169,13 +171,22 @@ define define_curve
 
 LIBCOMPONENTS += $$(BUILD_OBJ)/$(1)/decaf.o $$(BUILD_OBJ)/$(1)/crypto.o $$(BUILD_OBJ)/$(1)/decaf_tables.o
 PER_OBJ_DIRS += $$(BUILD_OBJ)/$(1)
-HEADERS_OF_$(1) = $$(HEADERS_OF_$(2))
+GLOBAL_HEADERS_OF_$(1) = $(BUILD_INC)/decaf/decaf_$(3).h $(BUILD_INC)/decaf/decaf_$(3).hxx \
+		$(BUILD_INC)/decaf/crypto_$(3).h $(BUILD_INC)/decaf/crypto_$(3).hxx
+HEADERS_OF_$(1) = $$(HEADERS_OF_$(2)) $$(GLOBAL_HEADERS_OF_$(1))
+HEADERS += $$(GLOBAL_HEADERS_OF_$(1))
 
 $$(BUILD_C)/$(1)/%.c: src/per_curve/%.tmpl.c src/gen_headers/* $$(HEADERS_OF_$(2))
 	python -B src/gen_headers/template.py --per=curve --item=$(1) --guard=$(1)/`basename $$@` -o $$@ $$<
 	
 $$(BUILD_H)/$(1)/%.h: src/per_curve/%.tmpl.h src/gen_headers/* $$(HEADERS_OF_$(2))
 	python -B src/gen_headers/template.py --per=curve --item=$(1) --guard=$(1)/`basename $$@` -o $$@ $$<
+	
+$$(BUILD_INC)/decaf/decaf_$(3).%: src/per_curve/decaf.tmpl.% src/gen_headers/* $$(HEADERS_OF_$(2))
+	python -B src/gen_headers/template.py --per=curve --item=$(1) --guard=$$(@:$(BUILD_INC)/%=%) -o $$@ $$<
+	
+$$(BUILD_INC)/decaf/crypto_$(3).%: src/per_curve/crypto.tmpl.% src/gen_headers/* $$(HEADERS_OF_$(2))
+	python -B src/gen_headers/template.py --per=curve --item=$(1) --guard=$$(@:$(BUILD_INC)/%=%) -o $$@ $$<
 
 $$(BUILD_IBIN)/decaf_gen_tables_$(1): $$(BUILD_OBJ)/$(1)/decaf_gen_tables.o \
 		$$(BUILD_OBJ)/$(1)/decaf.o $$(BUILD_OBJ)/utils.o \
@@ -200,9 +211,9 @@ endef
 ################################################################
 # call code above to generate curves and fields
 $(eval $(call define_field,p25519,arch_x86_64))
-$(eval $(call define_curve,curve25519,p25519))
+$(eval $(call define_curve,curve25519,p25519,255))
 $(eval $(call define_field,p448,arch_x86_64))
-$(eval $(call define_curve,ed448goldilocks,p448))
+$(eval $(call define_curve,ed448goldilocks,p448,448))
 
 # The shakesum utility is in the public bin directory.
 $(BUILD_BIN)/shakesum: $(BUILD_OBJ)/shakesum.o $(BUILD_OBJ)/shake.o $(BUILD_OBJ)/utils.o
diff --git a/src/gen_headers/curve_data.py b/src/gen_headers/curve_data.py
index d9f26ec..17813b1 100644
--- a/src/gen_headers/curve_data.py
+++ b/src/gen_headers/curve_data.py
@@ -84,6 +84,7 @@ def ceil_log2(x):
         out += 1
     return out
 
+# TODO: reduce this because we can now have expressions.
 for field,data in field_data.iteritems():
     if "modulus" not in data:
         data["modulus"] = eval(data["gf_desc"].replace("^","**"))
diff --git a/src/gen_headers/gen_file.py b/src/gen_headers/gen_file.py
deleted file mode 100644
index ea4b5b3..0000000
--- a/src/gen_headers/gen_file.py
+++ /dev/null
@@ -1,50 +0,0 @@
-from curve_data import curve_data, field_data
-from textwrap import dedent
-
-def redoc(filename,doc,author):
-    doc = doc.replace("\n","\n * ")
-    doc = dedent("""
-        /**
-         * @file %(filename)s
-         * @author %(author)s
-         *
-         * @copyright
-         *   Copyright (c) 2015-2016 Cryptography Research, Inc.  \\n
-         *   Released under the MIT License.  See LICENSE.txt for license information.
-         *
-         * %(doc)s
-         *
-         * @warning This file was automatically generated in Python.
-         * Please do not edit it.
-         */""") % { "filename": filename, "doc": doc, "author" : author }
-    doc = doc.replace(" * \n", " *\n")
-    return doc[1:]
-
-gend_files = {}
-
-per_map = {"field":field_data, "curve":curve_data, "global":{"global":{}} }
-
-def gen_file(public,name,doc,code,per="global",author="Mike Hamburg"):
-    is_header = name.endswith(".h") or name.endswith(".hxx") or name.endswith(".h++")
-    
-    for curve,data in per_map[per].iteritems():
-        ns_name = name % data
-        
-        _,_,name_base = ns_name.rpartition("/")
-        header_guard = "__" + ns_name.replace(".","_").replace("/","_").upper() + "__"
-        
-        ns_doc = dedent(doc).strip().rstrip()
-        ns_doc = redoc(ns_name, ns_doc % data, author)
-        ns_code = code % data
-        ret = ns_doc + "\n"
-        
-        if is_header:
-            ns_code = dedent("""\n
-                #ifndef %(header_guard)s
-                #define %(header_guard)s 1
-                %(code)s
-                #endif /* %(header_guard)s */
-                """) % { "header_guard" : header_guard, "code": ns_code }
-        ret += ns_code[1:-1]
-        
-        gend_files[ns_name] = (public,ret)
diff --git a/src/gen_headers/main.py b/src/gen_headers/main.py
deleted file mode 100644
index 288041b..0000000
--- a/src/gen_headers/main.py
+++ /dev/null
@@ -1,119 +0,0 @@
-from gen_file import gen_file,gend_files
-
-import os
-import argparse
-import re
-
-parser = argparse.ArgumentParser(description='Generate Decaf headers and other such files.')
-parser.add_argument('--hpre', required = True, help = "Where to put the public header files")
-parser.add_argument('--ihpre', required = True, help = "Where to put the internal header files")
-parser.add_argument('--cpre', required = True, help = "Where to put the C/C++ implementation files")
-args = parser.parse_args()
-
-prefixes = { (True,"h") : args.hpre, (True,"hxx") : args.hpre, (False,"c") : args.cpre, (False,"h") : args.ihpre }
-
-from decaf_hxx import decaf_hxx
-from decaf_h import decaf_h
-from crypto_h import crypto_h
-from crypto_hxx import crypto_hxx
-from curve_data import curve_data
-
-root_hxx_code = "\n".join((
-    "#include <%s>" % name
-    for name in sorted(gend_files)
-    if re.match("^decaf/decaf_\d+.hxx$",name)
-))
-root_hxx_code += """
-
-namespace decaf {
-    template <template<typename Group> class Run>
-    void run_for_all_curves() {
-"""
-root_hxx_code += "\n".join((
-    "        Run<%s>::run();" % cd["cxx_ns"]
-    for cd in sorted(curve_data.values(), key=lambda x:x["c_ns"])
-))
-root_hxx_code += """
-    }
-}
-"""
-decaf_root_hxx = gen_file(
-    public = True,
-    per = "global",
-    name = "decaf.hxx",
-    doc = """@brief Decaf curve metaheader.""",
-    code = "\n"+root_hxx_code+"\n"
-)
-
-crypto_h_code = "\n".join((
-    "#include <%s>" % name
-    for name in sorted(gend_files)
-    if re.match("^decaf/crypto_\d+.h$",name)
-))
-crypto_h = gen_file(
-    public = True,
-    per = "global",
-    name = "decaf/crypto.h",
-    doc = """
-        Example Decaf crypto routines, metaheader.
-        @warning These are merely examples, though they ought to be secure.  But real
-        protocols will decide differently on magic numbers, formats, which items to
-        hash, etc.
-    """,
-    code = "\n"+crypto_h_code+"\n"
-)
-
-crypto_hxx_code = "\n".join((
-    "#include <%s>" % name
-    for name in sorted(gend_files)
-    if re.match("^decaf/crypto_\d+.hxx$",name)
-))
-crypto_hxx = gen_file(
-    public = True,
-    per = "global",
-    name = "decaf/crypto.hxx",
-    doc = """
-        Example Decaf crypto routines, C++, metaheader.
-        @warning These are merely examples, though they ought to be secure.  But real
-        protocols will decide differently on magic numbers, formats, which items to
-        hash, etc.
-    """,
-    code = "\n"+crypto_hxx_code+"\n"
-)
-
-root_h_code = "\n".join((
-    "#include <%s>" % name
-    for name in sorted(gend_files)
-    if re.match("^decaf/decaf_\d+.h$",name)
-))
-decaf_root_hxx = gen_file(
-    public = True,
-    per = "global",
-    name = "decaf.h",
-    doc = """
-        Master header for Decaf library.
-        
-        The Decaf library implements cryptographic operations on a elliptic curve
-        groups of prime order p.  It accomplishes this by using a twisted Edwards
-        curve (isogenous to Ed448-Goldilocks or Ed25519) and wiping out the cofactor.
-        
-        The formulas are all complete and have no special cases.  However, some
-        functions can fail.  For example, decoding functions can fail because not
-        every string is the encoding of a valid group element.
-        
-        The formulas contain no data-dependent branches, timing or memory accesses,
-        except for decaf_XXX_base_double_scalarmul_non_secret.
-    """,
-    code = "\n"+root_h_code+"\n"
-)
-
-
-for name,(public,code) in gend_files.iteritems():        
-    _,_,name_suffix = name.rpartition(".")
-    prefix = prefixes[(public,name_suffix)]
-    if not os.path.exists(os.path.dirname(prefix + "/" + name)):
-        os.makedirs(os.path.dirname(prefix + "/" + name))
-    with open(prefix + "/" + name,"w") as f:
-        f.write(code + "\n")
-    
-    
\ No newline at end of file
diff --git a/src/gen_headers/template.py b/src/gen_headers/template.py
index ba56cc8..0f9e380 100644
--- a/src/gen_headers/template.py
+++ b/src/gen_headers/template.py
@@ -13,7 +13,7 @@ parser.add_argument('--guard', required = False, default = None, help = "header
 parser.add_argument('files', metavar='file', type=str, nargs='+', help='a list of files to fill')
 args = parser.parse_args()
 
-per_map = {"field":field_data, "curve":curve_data, "global":{"global":{}} }
+per_map = {"field":field_data, "curve":curve_data, "global":{"global":{"field":field_data,"curve":curve_data} }}
 
 def redoc(filename,doc,author):
     doc = doc.replace("\n","\n * ")
diff --git a/src/gen_headers/crypto_h.py b/src/per_curve/crypto.tmpl.h
similarity index 63%
rename from src/gen_headers/crypto_h.py
rename to src/per_curve/crypto.tmpl.h
index 8d5df60..39cab3e 100644
--- a/src/gen_headers/crypto_h.py
+++ b/src/per_curve/crypto.tmpl.h
@@ -1,17 +1,12 @@
-from gen_file import gen_file
+/**
+ * Example Decaf crypto routines.
+ * @warning These are merely examples, though they ought to be secure.  But real
+ * protocols will decide differently on magic numbers, formats, which items to
+ * hash, etc.
+ * @warning Experimental!  The names, parameter orders etc are likely to change.
+ */
 
-crypto_h = gen_file(
-    public = True,
-    per = "curve",
-    name = "decaf/crypto_%(shortname)s.h",
-    doc = """
-        Example Decaf crypto routines.
-        @warning These are merely examples, though they ought to be secure.  But real
-        protocols will decide differently on magic numbers, formats, which items to
-        hash, etc.
-        @warning Experimental!  The names, parameter orders etc are likely to change.
-    """, code = """
-#include <decaf/%(c_ns)s.h>
+#include <decaf/$(c_ns).h>
 #include <decaf/strobe.h>
 
 #ifdef __cplusplus
@@ -19,48 +14,48 @@ extern "C" {
 #endif
 
 /** Number of bytes for a symmetric key (expanded to full key) */
-#define %(C_NS)s_SYMMETRIC_KEY_BYTES 32
+#define $(C_NS)_SYMMETRIC_KEY_BYTES 32
 
 /** A symmetric key, the compressed point of a private key. */
-typedef unsigned char %(c_ns)s_symmetric_key_t[%(C_NS)s_SYMMETRIC_KEY_BYTES];
+typedef unsigned char $(c_ns)_symmetric_key_t[$(C_NS)_SYMMETRIC_KEY_BYTES];
 
 /** An encoded public key. */
-typedef unsigned char %(c_ns)s_public_key_t[%(C_NS)s_SER_BYTES];
+typedef unsigned char $(c_ns)_public_key_t[$(C_NS)_SER_BYTES];
 
 /** A signature. */
-typedef unsigned char %(c_ns)s_signature_t[%(C_NS)s_SER_BYTES + %(C_NS)s_SCALAR_BYTES];
+typedef unsigned char $(c_ns)_signature_t[$(C_NS)_SER_BYTES + $(C_NS)_SCALAR_BYTES];
 
 typedef struct {
     /** @cond internal */
     /** The symmetric key from which everything is expanded */
-    %(c_ns)s_symmetric_key_t sym;
+    $(c_ns)_symmetric_key_t sym;
     
     /** The scalar x */
-    %(c_ns)s_scalar_t secret_scalar;
+    $(c_ns)_scalar_t secret_scalar;
     
     /** x*Base */
-    %(c_ns)s_public_key_t pub;
+    $(c_ns)_public_key_t pub;
     /** @endcond */
 } /** Private key structure for pointers. */
-  %(c_ns)s_private_key_s,
+  $(c_ns)_private_key_s,
   /** A private key (gmp array[1] style). */
-  %(c_ns)s_private_key_t[1];
+  $(c_ns)_private_key_t[1];
     
 /**
  * Derive a key from its compressed form.
  * @param [out] priv The derived private key.
  * @param [in] proto The compressed or proto-key, which must be 32 random bytes.
  */
-void %(c_ns)s_derive_private_key (
-    %(c_ns)s_private_key_t priv,
-    const %(c_ns)s_symmetric_key_t proto
+void $(c_ns)_derive_private_key (
+    $(c_ns)_private_key_t priv,
+    const $(c_ns)_symmetric_key_t proto
 ) NONNULL2 API_VIS;
 
 /**
  * Destroy a private key.
  */
-void %(c_ns)s_destroy_private_key (
-    %(c_ns)s_private_key_t priv
+void $(c_ns)_destroy_private_key (
+    $(c_ns)_private_key_t priv
 ) NONNULL1 API_VIS;
 
 /**
@@ -68,9 +63,9 @@ void %(c_ns)s_destroy_private_key (
  * @param [out] pub The extracted private key.
  * @param [in] priv The private key.
  */
-void %(c_ns)s_private_to_public (
-    %(c_ns)s_public_key_t pub,
-    const %(c_ns)s_private_key_t priv
+void $(c_ns)_private_to_public (
+    $(c_ns)_public_key_t pub,
+    const $(c_ns)_private_key_t priv
 ) NONNULL2 API_VIS;
     
 /**
@@ -89,11 +84,11 @@ void %(c_ns)s_private_to_public (
  * @retval DECAF_FAILURE Key exchange failed.
  */
 decaf_error_t
-%(c_ns)s_shared_secret (
+$(c_ns)_shared_secret (
     uint8_t *shared,
     size_t shared_bytes,
-    const %(c_ns)s_private_key_t my_privkey,
-    const %(c_ns)s_public_key_t your_pubkey,
+    const $(c_ns)_private_key_t my_privkey,
+    const $(c_ns)_public_key_t your_pubkey,
     int me_first
 ) NONNULL134 WARN_UNUSED API_VIS;
    
@@ -105,10 +100,10 @@ decaf_error_t
  * @param [in] strobe A STROBE context with the message.
  */ 
 void
-%(c_ns)s_sign_strobe (
+$(c_ns)_sign_strobe (
     keccak_strobe_t strobe,
-    %(c_ns)s_signature_t sig,
-    const %(c_ns)s_private_key_t priv
+    $(c_ns)_signature_t sig,
+    const $(c_ns)_private_key_t priv
 ) NONNULL3 API_VIS;
 
 /**
@@ -120,9 +115,9 @@ void
  * @param [in] message_len The message's length.
  */ 
 void
-%(c_ns)s_sign (
-    %(c_ns)s_signature_t sig,
-    const %(c_ns)s_private_key_t priv,
+$(c_ns)_sign (
+    $(c_ns)_signature_t sig,
+    const $(c_ns)_private_key_t priv,
     const unsigned char *message,
     size_t message_len
 ) NONNULL3 API_VIS;
@@ -138,10 +133,10 @@ void
  * @return DECAF_FAILURE The signature did not verify successfully.
  */    
 decaf_error_t
-%(c_ns)s_verify_strobe (
+$(c_ns)_verify_strobe (
     keccak_strobe_t strobe,
-    const %(c_ns)s_signature_t sig,
-    const %(c_ns)s_public_key_t pub
+    const $(c_ns)_signature_t sig,
+    const $(c_ns)_public_key_t pub
 ) NONNULL3 API_VIS WARN_UNUSED;
 
 /**
@@ -156,9 +151,9 @@ decaf_error_t
  * @return DECAF_FAILURE The signature did not verify successfully.
  */    
 decaf_error_t
-%(c_ns)s_verify (
-    const %(c_ns)s_signature_t sig,
-    const %(c_ns)s_public_key_t pub,
+$(c_ns)_verify (
+    const $(c_ns)_signature_t sig,
+    const $(c_ns)_public_key_t pub,
     const unsigned char *message,
     size_t message_len
 ) NONNULL3 API_VIS WARN_UNUSED;
@@ -166,4 +161,3 @@ decaf_error_t
 #ifdef __cplusplus
 } /* extern "C" */
 #endif
-""")
diff --git a/src/gen_headers/crypto_hxx.py b/src/per_curve/crypto.tmpl.hxx
similarity index 63%
rename from src/gen_headers/crypto_hxx.py
rename to src/per_curve/crypto.tmpl.hxx
index 7a0bf21..bc37c23 100644
--- a/src/gen_headers/crypto_hxx.py
+++ b/src/per_curve/crypto.tmpl.hxx
@@ -1,17 +1,12 @@
-from gen_file import gen_file
+/*
+ * Example Decaf cyrpto routines, C++ wrapper.
+ * @warning These are merely examples, though they ought to be secure.  But real
+ * protocols will decide differently on magic numbers, formats, which items to
+ * hash, etc.
+ * @warning Experimental!  The names, parameter orders etc are likely to change.
+ */
 
-crypto_hxx = gen_file(
-    public = True,
-    per = "curve",
-    name = "decaf/crypto_%(shortname)s.hxx",
-    doc = """
-        Example Decaf cyrpto routines, C++ wrapper.
-        @warning These are merely examples, though they ought to be secure.  But real
-        protocols will decide differently on magic numbers, formats, which items to
-        hash, etc.
-        @warning Experimental!  The names, parameter orders etc are likely to change.
-    """, code = """
-#include <decaf.hxx>
+#include <decaf/decaf_$(gf_bits).hxx>
 #include <decaf/shake.hxx>
 #include <decaf/strobe.hxx>
 
@@ -31,21 +26,21 @@ template <typename Group> class PublicKey;
 /** A private key for crypto over some Group */
 template <typename Group> class PrivateKey;
 
-/** A public key for crypto over %(name)s */
-template<> class PublicKey<%(cxx_ns)s>
-  : public Serializable< PublicKey<%(cxx_ns)s> > {
+/** A public key for crypto over $(name) */
+template<> class PublicKey<$(cxx_ns)>
+  : public Serializable< PublicKey<$(cxx_ns)> > {
 private:
 /** @cond internal */
-    typedef %(c_ns)s_public_key_t Wrapped;
+    typedef $(c_ns)_public_key_t Wrapped;
     Wrapped wrapped;
     template<class Group> friend class PrivateKey;
 /** @endcond */
 public:
     /** Underlying group */
-    typedef %(cxx_ns)s Group;
+    typedef $(cxx_ns) Group;
     
     /** Signature size. */
-    static const size_t SIG_BYTES = sizeof(%(c_ns)s_signature_t);
+    static const size_t SIG_BYTES = sizeof($(c_ns)_signature_t);
     
     /** Serialization size. */
     static const size_t SER_BYTES = sizeof(Wrapped);
@@ -56,7 +51,7 @@ public:
     }
     
     /** Read a private key from a string*/
-    inline explicit PublicKey(const PrivateKey<%(cxx_ns)s> &b) NOEXCEPT;
+    inline explicit PublicKey(const PrivateKey<$(cxx_ns)> &b) NOEXCEPT;
     
     /** Create but don't initialize */
     inline explicit PublicKey(const NOINIT&) NOEXCEPT { }
@@ -74,7 +69,7 @@ public:
         const Block &message,
         const FixedBlock<SIG_BYTES> &sig
     ) const throw(CryptoException) {
-        if (DECAF_SUCCESS != %(c_ns)s_verify(sig.data(),wrapped,message.data(),message.size())) {
+        if (DECAF_SUCCESS != $(c_ns)_verify(sig.data(),wrapped,message.data(),message.size())) {
             throw(CryptoException());
         }
     }
@@ -84,33 +79,33 @@ public:
         Strobe &context,
         const FixedBlock<SIG_BYTES> &sig
     ) const throw(CryptoException) {
-        if (DECAF_SUCCESS != %(c_ns)s_verify_strobe(context.wrapped,sig.data(),wrapped)) {
+        if (DECAF_SUCCESS != $(c_ns)_verify_strobe(context.wrapped,sig.data(),wrapped)) {
             throw(CryptoException());
         }
     }
 };
 
-/** A private key for crypto over %(name)s */
-template<> class PrivateKey<%(cxx_ns)s>
-  : public Serializable< PrivateKey<%(cxx_ns)s> > {
+/** A private key for crypto over $(name) */
+template<> class PrivateKey<$(cxx_ns)>
+  : public Serializable< PrivateKey<$(cxx_ns)> > {
 private:
 /** @cond internal */
-    typedef %(c_ns)s_private_key_t Wrapped;
+    typedef $(c_ns)_private_key_t Wrapped;
     Wrapped wrapped;
     template<class Group> friend class PublicKey;
 /** @endcond */
 public:
     /** Underlying group */
-    typedef %(cxx_ns)s Group;
+    typedef $(cxx_ns) Group;
     
     /** Signature size. */
-    static const size_t SIG_BYTES = sizeof(%(c_ns)s_signature_t);
+    static const size_t SIG_BYTES = sizeof($(c_ns)_signature_t);
     
     /** Serialization size. */
     static const size_t SER_BYTES = sizeof(Wrapped);
     
     /** Compressed size. */
-    static const size_t SYM_BYTES = %(C_NS)s_SYMMETRIC_KEY_BYTES;
+    static const size_t SYM_BYTES = $(C_NS)_SYMMETRIC_KEY_BYTES;
     
     /** Create but don't initialize */
     inline explicit PrivateKey(const NOINIT&) NOEXCEPT { }
@@ -122,18 +117,18 @@ public:
     
     /** Read a private key from a string*/
     inline explicit PrivateKey(const FixedBlock<SYM_BYTES> &b) NOEXCEPT {
-        %(c_ns)s_derive_private_key(wrapped, b.data());
+        $(c_ns)_derive_private_key(wrapped, b.data());
     }
     
     /** Create at random */
     inline explicit PrivateKey(Rng &r) NOEXCEPT {
         FixedArrayBuffer<SYM_BYTES> tmp(r);
-        %(c_ns)s_derive_private_key(wrapped, tmp.data());
+        $(c_ns)_derive_private_key(wrapped, tmp.data());
     }
     
     /** Secure destructor */
     inline ~PrivateKey() NOEXCEPT {
-        %(c_ns)s_destroy_private_key(wrapped);
+        $(c_ns)_destroy_private_key(wrapped);
     }
     
     /** Serialization size. */
@@ -152,18 +147,18 @@ public:
     }
     
     /** Get the public key */
-    inline PublicKey<%(cxx_ns)s> pub() const NOEXCEPT {
-        PublicKey<%(cxx_ns)s> ret(*this); return ret;
+    inline PublicKey<$(cxx_ns)> pub() const NOEXCEPT {
+        PublicKey<$(cxx_ns)> ret(*this); return ret;
     }
     
     /** Derive a shared secret */
     inline SecureBuffer sharedSecret(
-        const PublicKey<%(cxx_ns)s> &pub,
+        const PublicKey<$(cxx_ns)> &pub,
         size_t bytes,
         bool me_first
     ) const throw(CryptoException,std::bad_alloc) {
         SecureBuffer ret(bytes);
-        if (DECAF_SUCCESS != %(c_ns)s_shared_secret(ret.data(),bytes,wrapped,pub.wrapped,me_first)) {
+        if (DECAF_SUCCESS != $(c_ns)_shared_secret(ret.data(),bytes,wrapped,pub.wrapped,me_first)) {
             throw(CryptoException());
         }
         return ret;
@@ -173,33 +168,32 @@ public:
     inline decaf_error_t __attribute__((warn_unused_result))
     sharedSecretNoexcept(
         Buffer ret,
-        const PublicKey<%(cxx_ns)s> &pub,
+        const PublicKey<$(cxx_ns)> &pub,
         bool me_first
     ) const NOEXCEPT {
-        return %(c_ns)s_shared_secret(ret.data(),ret.size(),wrapped,pub.wrapped,me_first);
+        return $(c_ns)_shared_secret(ret.data(),ret.size(),wrapped,pub.wrapped,me_first);
     }
 
     /** Sign a message. */ 
     inline SecureBuffer sign(const Block &message) const {
         SecureBuffer sig(SIG_BYTES);
-        %(c_ns)s_sign(sig.data(), wrapped, message.data(), message.size());
+        $(c_ns)_sign(sig.data(), wrapped, message.data(), message.size());
         return sig;
     }
 
     /** Sign a message. */ 
     inline SecureBuffer verify(Strobe &context) const {
         SecureBuffer sig(SIG_BYTES);
-        %(c_ns)s_sign_strobe(context.wrapped, sig.data(), wrapped);
+        $(c_ns)_sign_strobe(context.wrapped, sig.data(), wrapped);
         return sig;
     }
 };
 
 /** @cond internal */
-PublicKey<%(cxx_ns)s>::PublicKey(const PrivateKey<%(cxx_ns)s> &b) NOEXCEPT {
-    %(c_ns)s_private_to_public(wrapped,b.wrapped);
+PublicKey<$(cxx_ns)>::PublicKey(const PrivateKey<$(cxx_ns)> &b) NOEXCEPT {
+    $(c_ns)_private_to_public(wrapped,b.wrapped);
 }
 /** @endcond */
 
 #undef NOEXCEPT
-} /* namespace decaf */
-""")
\ No newline at end of file
+} /* namespace decaf */
\ No newline at end of file
diff --git a/src/per_curve/decaf.tmpl.c b/src/per_curve/decaf.tmpl.c
index c123d94..dee9531 100644
--- a/src/per_curve/decaf.tmpl.c
+++ b/src/per_curve/decaf.tmpl.c
@@ -19,7 +19,7 @@
 #define IMAGINE_TWIST $(imagine_twist)
 #define COFACTOR $(cofactor)
 
-/** Comb config: number of combs, n, t, s. */
+/* Comb config: number of combs, n, t, s. */
 #define COMBS_N $(combs.n)
 #define COMBS_T $(combs.t)
 #define COMBS_S $(combs.s)
@@ -39,9 +39,7 @@ static const scalar_t sc_p = {{{
 }}};
 static const decaf_word_t MONTGOMERY_FACTOR = (decaf_word_t)0x$("%x" % pow(-q,2**64-1,2**64))ull;
 
-const uint8_t API_NS(x_base_point)[SER_BYTES] /* TODO */ = {
-    $(ser(mont_base,8))
-};
+const uint8_t API_NS(x_base_point)[SER_BYTES] = { $(ser(mont_base,8)) };
 
 #if COFACTOR==8
     static const gf SQRT_ONE_MINUS_D = {FIELD_LITERAL(
diff --git a/src/gen_headers/decaf_h.py b/src/per_curve/decaf.tmpl.h
similarity index 66%
rename from src/gen_headers/decaf_h.py
rename to src/per_curve/decaf.tmpl.h
index ceddec0..fdf1dae 100644
--- a/src/gen_headers/decaf_h.py
+++ b/src/per_curve/decaf.tmpl.h
@@ -1,11 +1,5 @@
-from gen_file import gen_file
-
-decaf_h = gen_file(
-    public = True,
-    per = "curve",
-    name = "decaf/%(c_ns)s.h",
-    doc = """@brief A group of prime order p, based on %(iso_to)s.""",
-    code = """
+/** @brief A group of prime order p, based on $(iso_to). */
+
 #include <decaf/common.h>
 
 #ifdef __cplusplus
@@ -13,78 +7,78 @@ extern "C" {
 #endif
 
 /** @cond internal */
-#define %(C_NS)s_SCALAR_LIMBS ((%(scalar_bits)d-1)/DECAF_WORD_BITS+1)
+#define $(C_NS)_SCALAR_LIMBS (($(scalar_bits)-1)/DECAF_WORD_BITS+1)
 /** @endcond */
 
 /** The number of bits in a scalar */
-#define %(C_NS)s_SCALAR_BITS %(scalar_bits)d
+#define $(C_NS)_SCALAR_BITS $(scalar_bits)
 
 /** @cond internal */
-#ifndef __DECAF_%(gf_shortname)s_GF_DEFINED__
-#define __DECAF_%(gf_shortname)s_GF_DEFINED__ 1
+#ifndef __DECAF_$(gf_shortname)_GF_DEFINED__
+#define __DECAF_$(gf_shortname)_GF_DEFINED__ 1
 /** @brief Galois field element internal structure */
-typedef struct gf_%(gf_shortname)s_s {
-    decaf_word_t limb[%(gf_impl_bits)d/DECAF_WORD_BITS];
-} __attribute__((aligned(32))) gf_%(gf_shortname)s_s, gf_%(gf_shortname)s_t[1];
-#endif /* __DECAF_%(gf_shortname)s_GF_DEFINED__ */
+typedef struct gf_$(gf_shortname)_s {
+    decaf_word_t limb[$(gf_impl_bits)/DECAF_WORD_BITS];
+} __attribute__((aligned(32))) gf_$(gf_shortname)_s, gf_$(gf_shortname)_t[1];
+#endif /* __DECAF_$(gf_shortname)_GF_DEFINED__ */
 /** @endcond */
 
 /** Number of bytes in a serialized point. */
-#define %(C_NS)s_SER_BYTES %(ser_bytes)d
+#define $(C_NS)_SER_BYTES $(ser_bytes)
 
 /** Number of bytes in a serialized scalar. */
-#define %(C_NS)s_SCALAR_BYTES %(scalar_ser_bytes)d
+#define $(C_NS)_SCALAR_BYTES $(scalar_ser_bytes)
 
-/** Number of bytes in an x%(gf_shortname)s public key */
-#define X%(gf_shortname)s_PUBLIC_BYTES %(x_pub_bytes)d
+/** Number of bytes in an x$(gf_shortname) public key */
+#define X$(gf_shortname)_PUBLIC_BYTES $(x_pub_bytes)
 
-/** Number of bytes in an x%(gf_shortname)s private key */
-#define X%(gf_shortname)s_PRIVATE_BYTES %(x_priv_bytes)d
+/** Number of bytes in an x$(gf_shortname) private key */
+#define X$(gf_shortname)_PRIVATE_BYTES $(x_priv_bytes)
 
 /** Twisted Edwards extended homogeneous coordinates */
-typedef struct %(c_ns)s_point_s {
+typedef struct $(c_ns)_point_s {
     /** @cond internal */
-    gf_%(gf_shortname)s_t x,y,z,t;
+    gf_$(gf_shortname)_t x,y,z,t;
     /** @endcond */
-} %(c_ns)s_point_t[1];
+} $(c_ns)_point_t[1];
 
 /** Precomputed table based on a point.  Can be trivial implementation. */
-struct %(c_ns)s_precomputed_s;
+struct $(c_ns)_precomputed_s;
 
 /** Precomputed table based on a point.  Can be trivial implementation. */
-typedef struct %(c_ns)s_precomputed_s %(c_ns)s_precomputed_s; 
+typedef struct $(c_ns)_precomputed_s $(c_ns)_precomputed_s; 
 
 /** Size and alignment of precomputed point tables. */
-extern const size_t %(c_ns)s_sizeof_precomputed_s API_VIS, %(c_ns)s_alignof_precomputed_s API_VIS;
+extern const size_t $(c_ns)_sizeof_precomputed_s API_VIS, $(c_ns)_alignof_precomputed_s API_VIS;
 
 /** Scalar is stored packed, because we don't need the speed. */
-typedef struct %(c_ns)s_scalar_s {
+typedef struct $(c_ns)_scalar_s {
     /** @cond internal */
-    decaf_word_t limb[%(C_NS)s_SCALAR_LIMBS];
+    decaf_word_t limb[$(C_NS)_SCALAR_LIMBS];
     /** @endcond */
-} %(c_ns)s_scalar_t[1];
+} $(c_ns)_scalar_t[1];
 
 /** A scalar equal to 1. */
-extern const %(c_ns)s_scalar_t %(c_ns)s_scalar_one API_VIS;
+extern const $(c_ns)_scalar_t $(c_ns)_scalar_one API_VIS;
 
 /** A scalar equal to 0. */
-extern const %(c_ns)s_scalar_t %(c_ns)s_scalar_zero API_VIS;
+extern const $(c_ns)_scalar_t $(c_ns)_scalar_zero API_VIS;
 
 /** The identity point on the curve. */
-extern const %(c_ns)s_point_t %(c_ns)s_point_identity API_VIS;
+extern const $(c_ns)_point_t $(c_ns)_point_identity API_VIS;
 
 /** An arbitrarily chosen base point on the curve.
  * @warning TODO: this is subject to change.  It is currently
- * the preimage of the X%(gf_shortname)s base point.  Sometime
- * soon, we will merge and finalize support for X%(gf_shortname)s
- * and Ed%(gf_shortname)s integration.  This might make some
+ * the preimage of the X$(gf_shortname) base point.  Sometime
+ * soon, we will merge and finalize support for X$(gf_shortname)
+ * and Ed$(gf_shortname) integration.  This might make some
  * multiple of the current basepoint (eg twice it, or the cofactor
  * times it) more convenient API-wise, and trigger a changeover.
  */
-extern const %(c_ns)s_point_t %(c_ns)s_point_base API_VIS;
+extern const $(c_ns)_point_t $(c_ns)_point_base API_VIS;
 
 /** Precomputed table for the base point on the curve. */
-extern const struct %(c_ns)s_precomputed_s *%(c_ns)s_precomputed_base API_VIS;
+extern const struct $(c_ns)_precomputed_s *$(c_ns)_precomputed_base API_VIS;
 
 /**
  * @brief Read a scalar from wire format or from bytes.
@@ -96,9 +90,9 @@ extern const struct %(c_ns)s_precomputed_s *%(c_ns)s_precomputed_base API_VIS;
  * @retval DECAF_FAILURE The scalar was greater than the modulus,
  * and has been reduced modulo that modulus.
  */
-decaf_error_t %(c_ns)s_scalar_decode (
-    %(c_ns)s_scalar_t out,
-    const unsigned char ser[%(C_NS)s_SCALAR_BYTES]
+decaf_error_t $(c_ns)_scalar_decode (
+    $(c_ns)_scalar_t out,
+    const unsigned char ser[$(C_NS)_SCALAR_BYTES]
 ) API_VIS WARN_UNUSED NONNULL2 NOINLINE;
 
 /**
@@ -109,8 +103,8 @@ decaf_error_t %(c_ns)s_scalar_decode (
  * @param [in] ser_len Length of serialized form.
  * @param [out] out Deserialized form.
  */
-void %(c_ns)s_scalar_decode_long (
-    %(c_ns)s_scalar_t out,
+void $(c_ns)_scalar_decode_long (
+    $(c_ns)_scalar_t out,
     const unsigned char *ser,
     size_t ser_len
 ) API_VIS NONNULL2 NOINLINE;
@@ -121,9 +115,9 @@ void %(c_ns)s_scalar_decode_long (
  * @param [out] ser Serialized form of a scalar.
  * @param [in] s Deserialized scalar.
  */
-void %(c_ns)s_scalar_encode (
-    unsigned char ser[%(C_NS)s_SCALAR_BYTES],
-    const %(c_ns)s_scalar_t s
+void $(c_ns)_scalar_encode (
+    unsigned char ser[$(C_NS)_SCALAR_BYTES],
+    const $(c_ns)_scalar_t s
 ) API_VIS NONNULL2 NOINLINE NOINLINE;
         
 /**
@@ -132,10 +126,10 @@ void %(c_ns)s_scalar_encode (
  * @param [in] b Another scalar.
  * @param [out] out a+b.
  */
-void %(c_ns)s_scalar_add (
-    %(c_ns)s_scalar_t out,
-    const %(c_ns)s_scalar_t a,
-    const %(c_ns)s_scalar_t b
+void $(c_ns)_scalar_add (
+    $(c_ns)_scalar_t out,
+    const $(c_ns)_scalar_t a,
+    const $(c_ns)_scalar_t b
 ) API_VIS NONNULL3 NOINLINE;
 
 /**
@@ -145,9 +139,9 @@ void %(c_ns)s_scalar_add (
  * @retval DECAF_TRUE The scalars are equal.
  * @retval DECAF_FALSE The scalars are not equal.
  */    
-decaf_bool_t %(c_ns)s_scalar_eq (
-    const %(c_ns)s_scalar_t a,
-    const %(c_ns)s_scalar_t b
+decaf_bool_t $(c_ns)_scalar_eq (
+    const $(c_ns)_scalar_t a,
+    const $(c_ns)_scalar_t b
 ) API_VIS WARN_UNUSED NONNULL2 NOINLINE;
 
 /**
@@ -156,10 +150,10 @@ decaf_bool_t %(c_ns)s_scalar_eq (
  * @param [in] b Another scalar.
  * @param [out] out a-b.
  */  
-void %(c_ns)s_scalar_sub (
-    %(c_ns)s_scalar_t out,
-    const %(c_ns)s_scalar_t a,
-    const %(c_ns)s_scalar_t b
+void $(c_ns)_scalar_sub (
+    $(c_ns)_scalar_t out,
+    const $(c_ns)_scalar_t a,
+    const $(c_ns)_scalar_t b
 ) API_VIS NONNULL3 NOINLINE;
 
 /**
@@ -168,10 +162,10 @@ void %(c_ns)s_scalar_sub (
  * @param [in] b Another scalar.
  * @param [out] out a*b.
  */  
-void %(c_ns)s_scalar_mul (
-    %(c_ns)s_scalar_t out,
-    const %(c_ns)s_scalar_t a,
-    const %(c_ns)s_scalar_t b
+void $(c_ns)_scalar_mul (
+    $(c_ns)_scalar_t out,
+    const $(c_ns)_scalar_t a,
+    const $(c_ns)_scalar_t b
 ) API_VIS NONNULL3 NOINLINE;
 
 /**
@@ -180,9 +174,9 @@ void %(c_ns)s_scalar_mul (
  * @param [out] out 1/a.
  * @return DECAF_SUCCESS The input is nonzero.
  */  
-decaf_error_t %(c_ns)s_scalar_invert (
-    %(c_ns)s_scalar_t out,
-    const %(c_ns)s_scalar_t a
+decaf_error_t $(c_ns)_scalar_invert (
+    $(c_ns)_scalar_t out,
+    const $(c_ns)_scalar_t a
 ) API_VIS WARN_UNUSED NONNULL2 NOINLINE;
 
 /**
@@ -191,9 +185,9 @@ decaf_error_t %(c_ns)s_scalar_invert (
  * @param [in] a A scalar.
  * @param [out] out Will become a copy of a.
  */
-static inline void NONNULL2 %(c_ns)s_scalar_copy (
-    %(c_ns)s_scalar_t out,
-    const %(c_ns)s_scalar_t a
+static inline void NONNULL2 $(c_ns)_scalar_copy (
+    $(c_ns)_scalar_t out,
+    const $(c_ns)_scalar_t a
 ) {
     *out = *a;
 }
@@ -203,8 +197,8 @@ static inline void NONNULL2 %(c_ns)s_scalar_copy (
  * @param [in] a An integer.
  * @param [out] out Will become equal to a.
  */  
-void %(c_ns)s_scalar_set_unsigned (
-    %(c_ns)s_scalar_t out,
+void $(c_ns)_scalar_set_unsigned (
+    $(c_ns)_scalar_t out,
     uint64_t a
 ) API_VIS NONNULL1;
 
@@ -214,9 +208,9 @@ void %(c_ns)s_scalar_set_unsigned (
  * @param [out] ser The byte representation of the point.
  * @param [in] pt The point to encode.
  */
-void %(c_ns)s_point_encode (
-    uint8_t ser[%(C_NS)s_SER_BYTES],
-    const %(c_ns)s_point_t pt
+void $(c_ns)_point_encode (
+    uint8_t ser[$(C_NS)_SER_BYTES],
+    const $(c_ns)_point_t pt
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
@@ -233,9 +227,9 @@ void %(c_ns)s_point_encode (
  * @retval DECAF_FAILURE The decoding didn't succeed, because
  * ser does not represent a point.
  */
-decaf_error_t %(c_ns)s_point_decode (
-    %(c_ns)s_point_t pt,
-    const uint8_t ser[%(C_NS)s_SER_BYTES],
+decaf_error_t $(c_ns)_point_decode (
+    $(c_ns)_point_t pt,
+    const uint8_t ser[$(C_NS)_SER_BYTES],
     decaf_bool_t allow_identity
 ) API_VIS WARN_UNUSED NONNULL2 NOINLINE;
 
@@ -246,9 +240,9 @@ decaf_error_t %(c_ns)s_point_decode (
  * @param [out] a A copy of the point.
  * @param [in] b Any point.
  */
-static inline void NONNULL2 %(c_ns)s_point_copy (
-    %(c_ns)s_point_t a,
-    const %(c_ns)s_point_t b
+static inline void NONNULL2 $(c_ns)_point_copy (
+    $(c_ns)_point_t a,
+    const $(c_ns)_point_t b
 ) {
     *a=*b;
 }
@@ -262,9 +256,9 @@ static inline void NONNULL2 %(c_ns)s_point_copy (
  * @retval DECAF_TRUE The points are equal.
  * @retval DECAF_FALSE The points are not equal.
  */
-decaf_bool_t %(c_ns)s_point_eq (
-    const %(c_ns)s_point_t a,
-    const %(c_ns)s_point_t b
+decaf_bool_t $(c_ns)_point_eq (
+    const $(c_ns)_point_t a,
+    const $(c_ns)_point_t b
 ) API_VIS WARN_UNUSED NONNULL2 NOINLINE;
 
 /**
@@ -276,22 +270,22 @@ decaf_bool_t %(c_ns)s_point_eq (
  * @param [in] a An addend.
  * @param [in] b An addend.
  */
-void %(c_ns)s_point_add (
-    %(c_ns)s_point_t sum,
-    const %(c_ns)s_point_t a,
-    const %(c_ns)s_point_t b
+void $(c_ns)_point_add (
+    $(c_ns)_point_t sum,
+    const $(c_ns)_point_t a,
+    const $(c_ns)_point_t b
 ) API_VIS NONNULL3;
 
 /**
  * @brief Double a point.  Equivalent to
- * %(c_ns)s_point_add(two_a,a,a), but potentially faster.
+ * $(c_ns)_point_add(two_a,a,a), but potentially faster.
  *
  * @param [out] two_a The sum a+a.
  * @param [in] a A point.
  */
-void %(c_ns)s_point_double (
-    %(c_ns)s_point_t two_a,
-    const %(c_ns)s_point_t a
+void $(c_ns)_point_double (
+    $(c_ns)_point_t two_a,
+    const $(c_ns)_point_t a
 ) API_VIS NONNULL2;
 
 /**
@@ -303,10 +297,10 @@ void %(c_ns)s_point_double (
  * @param [in] a The minuend.
  * @param [in] b The subtrahend.
  */
-void %(c_ns)s_point_sub (
-    %(c_ns)s_point_t diff,
-    const %(c_ns)s_point_t a,
-    const %(c_ns)s_point_t b
+void $(c_ns)_point_sub (
+    $(c_ns)_point_t diff,
+    const $(c_ns)_point_t a,
+    const $(c_ns)_point_t b
 ) API_VIS NONNULL3;
     
 /**
@@ -316,9 +310,9 @@ void %(c_ns)s_point_sub (
  * @param [out] nega The negated input point
  * @param [in] a The input point.
  */
-void %(c_ns)s_point_negate (
-   %(c_ns)s_point_t nega,
-   const %(c_ns)s_point_t a
+void $(c_ns)_point_negate (
+   $(c_ns)_point_t nega,
+   const $(c_ns)_point_t a
 ) API_VIS NONNULL2;
 
 /**
@@ -328,10 +322,10 @@ void %(c_ns)s_point_negate (
  * @param [in] base The point to be scaled.
  * @param [in] scalar The scalar to multiply by.
  */
-void %(c_ns)s_point_scalarmul (
-    %(c_ns)s_point_t scaled,
-    const %(c_ns)s_point_t base,
-    const %(c_ns)s_scalar_t scalar
+void $(c_ns)_point_scalarmul (
+    $(c_ns)_point_t scaled,
+    const $(c_ns)_point_t base,
+    const $(c_ns)_scalar_t scalar
 ) API_VIS NONNULL3 NOINLINE;
 
 /**
@@ -351,10 +345,10 @@ void %(c_ns)s_point_scalarmul (
  * @retval DECAF_FAILURE The scalarmul didn't succeed, because
  * base does not represent a point.
  */
-decaf_error_t %(c_ns)s_direct_scalarmul (
-    uint8_t scaled[%(C_NS)s_SER_BYTES],
-    const uint8_t base[%(C_NS)s_SER_BYTES],
-    const %(c_ns)s_scalar_t scalar,
+decaf_error_t $(c_ns)_direct_scalarmul (
+    uint8_t scaled[$(C_NS)_SER_BYTES],
+    const uint8_t base[$(C_NS)_SER_BYTES],
+    const $(c_ns)_scalar_t scalar,
     decaf_bool_t allow_identity,
     decaf_bool_t short_circuit
 ) API_VIS NONNULL3 WARN_UNUSED NOINLINE;
@@ -371,14 +365,14 @@ decaf_error_t %(c_ns)s_direct_scalarmul (
  * @retval DECAF_FAILURE The scalarmul didn't succeed, because the base
  * point is in a small subgroup.
  */
-decaf_error_t %(c_ns)s_x_direct_scalarmul ( /* TODO: rename? */
-    uint8_t out[X%(gf_shortname)s_PUBLIC_BYTES],
-    const uint8_t base[X%(gf_shortname)s_PUBLIC_BYTES],
-    const uint8_t scalar[X%(gf_shortname)s_PRIVATE_BYTES]
+decaf_error_t $(c_ns)_x_direct_scalarmul ( /* TODO: rename? */
+    uint8_t out[X$(gf_shortname)_PUBLIC_BYTES],
+    const uint8_t base[X$(gf_shortname)_PUBLIC_BYTES],
+    const uint8_t scalar[X$(gf_shortname)_PRIVATE_BYTES]
 ) API_VIS NONNULL3 WARN_UNUSED NOINLINE;
 
-/** The base point for X%(gf_shortname)s Diffie-Hellman */
-extern const uint8_t %(c_ns)s_x_base_point[X%(gf_shortname)s_PUBLIC_BYTES] API_VIS;
+/** The base point for X$(gf_shortname) Diffie-Hellman */
+extern const uint8_t $(c_ns)_x_base_point[X$(gf_shortname)_PUBLIC_BYTES] API_VIS;
 
 /**
  * @brief RFC 7748 Diffie-Hellman base point scalarmul.  This function uses
@@ -387,9 +381,9 @@ extern const uint8_t %(c_ns)s_x_base_point[X%(gf_shortname)s_PUBLIC_BYTES] API_V
  * @param [out] scaled The scaled point base*scalar
  * @param [in] scalar The scalar to multiply by.
  */
-void %(c_ns)s_x_base_scalarmul (
-    uint8_t out[X%(gf_shortname)s_PUBLIC_BYTES],
-    const uint8_t scalar[X%(gf_shortname)s_PRIVATE_BYTES]
+void $(c_ns)_x_base_scalarmul (
+    uint8_t out[X$(gf_shortname)_PUBLIC_BYTES],
+    const uint8_t scalar[X$(gf_shortname)_PRIVATE_BYTES]
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
@@ -401,9 +395,9 @@ void %(c_ns)s_x_base_scalarmul (
  * @param [out] a A precomputed table of multiples of the point.
  * @param [in] b Any point.
  */
-void %(c_ns)s_precompute (
-    %(c_ns)s_precomputed_s *a,
-    const %(c_ns)s_point_t b
+void $(c_ns)_precompute (
+    $(c_ns)_precomputed_s *a,
+    const $(c_ns)_point_t b
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
@@ -411,23 +405,23 @@ void %(c_ns)s_precompute (
  * scaled = scalar*base.
  * Some implementations do not include precomputed points; for
  * those implementations, this function is the same as
- * %(c_ns)s_point_scalarmul
+ * $(c_ns)_point_scalarmul
  *
  * @param [out] scaled The scaled point base*scalar
  * @param [in] base The point to be scaled.
  * @param [in] scalar The scalar to multiply by.
  */
-void %(c_ns)s_precomputed_scalarmul (
-    %(c_ns)s_point_t scaled,
-    const %(c_ns)s_precomputed_s *base,
-    const %(c_ns)s_scalar_t scalar
+void $(c_ns)_precomputed_scalarmul (
+    $(c_ns)_point_t scaled,
+    const $(c_ns)_precomputed_s *base,
+    const $(c_ns)_scalar_t scalar
 ) API_VIS NONNULL3 NOINLINE;
 
 /**
  * @brief Multiply two base points by two scalars:
  * scaled = scalar1*base1 + scalar2*base2.
  *
- * Equivalent to two calls to %(c_ns)s_point_scalarmul, but may be
+ * Equivalent to two calls to $(c_ns)_point_scalarmul, but may be
  * faster.
  *
  * @param [out] combo The linear combination scalar1*base1 + scalar2*base2.
@@ -436,12 +430,12 @@ void %(c_ns)s_precomputed_scalarmul (
  * @param [in] base2 A second point to be scaled.
  * @param [in] scalar2 A second scalar to multiply by.
  */
-void %(c_ns)s_point_double_scalarmul (
-    %(c_ns)s_point_t combo,
-    const %(c_ns)s_point_t base1,
-    const %(c_ns)s_scalar_t scalar1,
-    const %(c_ns)s_point_t base2,
-    const %(c_ns)s_scalar_t scalar2
+void $(c_ns)_point_double_scalarmul (
+    $(c_ns)_point_t combo,
+    const $(c_ns)_point_t base1,
+    const $(c_ns)_scalar_t scalar1,
+    const $(c_ns)_point_t base2,
+    const $(c_ns)_scalar_t scalar2
 ) API_VIS NONNULL5 NOINLINE;
     
 /**
@@ -450,7 +444,7 @@ void %(c_ns)s_point_double_scalarmul (
  * a1 = scalar1 * base
  * a2 = scalar2 * base
  *
- * Equivalent to two calls to %(c_ns)s_point_scalarmul, but may be
+ * Equivalent to two calls to $(c_ns)_point_scalarmul, but may be
  * faster.
  *
  * @param [out] a1 The first multiple.  It may be the same as the input point.
@@ -459,19 +453,19 @@ void %(c_ns)s_point_double_scalarmul (
  * @param [in] scalar1 A first scalar to multiply by.
  * @param [in] scalar2 A second scalar to multiply by.
  */
-void %(c_ns)s_point_dual_scalarmul (
-    %(c_ns)s_point_t a1,
-    %(c_ns)s_point_t a2,
-    const %(c_ns)s_point_t base1,
-    const %(c_ns)s_scalar_t scalar1,
-    const %(c_ns)s_scalar_t scalar2
+void $(c_ns)_point_dual_scalarmul (
+    $(c_ns)_point_t a1,
+    $(c_ns)_point_t a2,
+    const $(c_ns)_point_t base1,
+    const $(c_ns)_scalar_t scalar1,
+    const $(c_ns)_scalar_t scalar2
 ) API_VIS NONNULL5 NOINLINE;
 
 /**
  * @brief Multiply two base points by two scalars:
- * scaled = scalar1*%(c_ns)s_point_base + scalar2*base2.
+ * scaled = scalar1*$(c_ns)_point_base + scalar2*base2.
  *
- * Otherwise equivalent to %(c_ns)s_point_double_scalarmul, but may be
+ * Otherwise equivalent to $(c_ns)_point_double_scalarmul, but may be
  * faster at the expense of being variable time.
  *
  * @param [out] combo The linear combination scalar1*base + scalar2*base2.
@@ -482,11 +476,11 @@ void %(c_ns)s_point_dual_scalarmul (
  * @warning: This function takes variable time, and may leak the scalars
  * used.  It is designed for signature verification.
  */
-void %(c_ns)s_base_double_scalarmul_non_secret (
-    %(c_ns)s_point_t combo,
-    const %(c_ns)s_scalar_t scalar1,
-    const %(c_ns)s_point_t base2,
-    const %(c_ns)s_scalar_t scalar2
+void $(c_ns)_base_double_scalarmul_non_secret (
+    $(c_ns)_point_t combo,
+    const $(c_ns)_scalar_t scalar1,
+    const $(c_ns)_point_t base2,
+    const $(c_ns)_scalar_t scalar2
 ) API_VIS NONNULL4 NOINLINE;
 
 /**
@@ -498,10 +492,10 @@ void %(c_ns)s_base_double_scalarmul_non_secret (
  * @param [in] b Any point.
  * @param [in] pick_b If nonzero, choose point b.
  */
-void %(c_ns)s_point_cond_sel (
-    %(c_ns)s_point_t out,
-    const %(c_ns)s_point_t a,
-    const %(c_ns)s_point_t b,
+void $(c_ns)_point_cond_sel (
+    $(c_ns)_point_t out,
+    const $(c_ns)_point_t a,
+    const $(c_ns)_point_t b,
     decaf_word_t pick_b
 ) API_VIS NONNULL3 NOINLINE;
 
@@ -514,10 +508,10 @@ void %(c_ns)s_point_cond_sel (
  * @param [in] b Any scalar.
  * @param [in] pick_b If nonzero, choose scalar b.
  */
-void %(c_ns)s_scalar_cond_sel (
-    %(c_ns)s_scalar_t out,
-    const %(c_ns)s_scalar_t a,
-    const %(c_ns)s_scalar_t b,
+void $(c_ns)_scalar_cond_sel (
+    $(c_ns)_scalar_t out,
+    const $(c_ns)_scalar_t a,
+    const $(c_ns)_scalar_t b,
     decaf_word_t pick_b
 ) API_VIS NONNULL3 NOINLINE;
 
@@ -528,8 +522,8 @@ void %(c_ns)s_scalar_cond_sel (
  * @retval DECAF_TRUE The point is valid.
  * @retval DECAF_FALSE The point is invalid.
  */
-decaf_bool_t %(c_ns)s_point_valid (
-    const %(c_ns)s_point_t toTest
+decaf_bool_t $(c_ns)_point_valid (
+    const $(c_ns)_point_t toTest
 ) API_VIS WARN_UNUSED NONNULL1 NOINLINE;
 
 /**
@@ -539,9 +533,9 @@ decaf_bool_t %(c_ns)s_point_valid (
  * @param [out] q The point to torque.
  * @param [in] p The point to torque.
  */
-void %(c_ns)s_point_debugging_torque (
-    %(c_ns)s_point_t q,
-    const %(c_ns)s_point_t p
+void $(c_ns)_point_debugging_torque (
+    $(c_ns)_point_t q,
+    const $(c_ns)_point_t p
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
@@ -553,10 +547,10 @@ void %(c_ns)s_point_debugging_torque (
  * @param [in] p The point to scale.
  * @param [in] factor Serialized GF factor to scale.
  */
-void %(c_ns)s_point_debugging_pscale (
-    %(c_ns)s_point_t q,
-    const %(c_ns)s_point_t p,
-    const unsigned char factor[%(C_NS)s_SER_BYTES]
+void $(c_ns)_point_debugging_pscale (
+    $(c_ns)_point_t q,
+    const $(c_ns)_point_t p,
+    const unsigned char factor[$(C_NS)_SER_BYTES]
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
@@ -564,7 +558,7 @@ void %(c_ns)s_point_debugging_pscale (
  *
  * Call this function with the output of a hash to make a hash to the curve.
  *
- * This function runs Elligator2 on the %(c_ns)s Jacobi quartic model.  It then
+ * This function runs Elligator2 on the $(c_ns) Jacobi quartic model.  It then
  * uses the isogeny to put the result in twisted Edwards form.  As a result,
  * it is safe (cannot produce points of order 4), and would be compatible with
  * hypothetical other implementations of Decaf using a Montgomery or untwisted
@@ -588,29 +582,29 @@ void %(c_ns)s_point_debugging_pscale (
  * @param [out] pt The data hashed to the curve.
  */
 void
-%(c_ns)s_point_from_hash_nonuniform (
-    %(c_ns)s_point_t pt,
-    const unsigned char hashed_data[%(C_NS)s_SER_BYTES]
+$(c_ns)_point_from_hash_nonuniform (
+    $(c_ns)_point_t pt,
+    const unsigned char hashed_data[$(C_NS)_SER_BYTES]
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
  * @brief Indifferentiable hash function encoding to curve.
  *
- * Equivalent to calling %(c_ns)s_point_from_hash_nonuniform twice and adding.
+ * Equivalent to calling $(c_ns)_point_from_hash_nonuniform twice and adding.
  *
  * @param [in] hashed_data Output of some hash function.
  * @param [out] pt The data hashed to the curve.
  */ 
-void %(c_ns)s_point_from_hash_uniform (
-    %(c_ns)s_point_t pt,
-    const unsigned char hashed_data[2*%(C_NS)s_SER_BYTES]
+void $(c_ns)_point_from_hash_uniform (
+    $(c_ns)_point_t pt,
+    const unsigned char hashed_data[2*$(C_NS)_SER_BYTES]
 ) API_VIS NONNULL2 NOINLINE;
 
 /**
  * @brief Inverse of elligator-like hash to curve.
  *
  * This function writes to the buffer, to make it so that
- * %(c_ns)s_point_from_hash_nonuniform(buffer) = pt if
+ * $(c_ns)_point_from_hash_nonuniform(buffer) = pt if
  * possible.  Since there may be multiple preimages, the
  * "which" parameter chooses between them.  To ensure uniform
  * inverse sampling, this function succeeds or fails
@@ -625,9 +619,9 @@ void %(c_ns)s_point_from_hash_uniform (
  * @retval DECAF_FAILURE The inverse failed.
  */
 decaf_error_t
-%(c_ns)s_invert_elligator_nonuniform (
-    unsigned char recovered_hash[%(C_NS)s_SER_BYTES],
-    const %(c_ns)s_point_t pt,
+$(c_ns)_invert_elligator_nonuniform (
+    unsigned char recovered_hash[$(C_NS)_SER_BYTES],
+    const $(c_ns)_point_t pt,
     uint16_t which
 ) API_VIS NONNULL2 NOINLINE WARN_UNUSED;
 
@@ -635,7 +629,7 @@ decaf_error_t
  * @brief Inverse of elligator-like hash to curve.
  *
  * This function writes to the buffer, to make it so that
- * %(c_ns)s_point_from_hash_uniform(buffer) = pt if
+ * $(c_ns)_point_from_hash_uniform(buffer) = pt if
  * possible.  Since there may be multiple preimages, the
  * "which" parameter chooses between them.  To ensure uniform
  * inverse sampling, this function succeeds or fails
@@ -650,36 +644,34 @@ decaf_error_t
  * @retval DECAF_FAILURE The inverse failed.
  */
 decaf_error_t
-%(c_ns)s_invert_elligator_uniform (
-    unsigned char recovered_hash[2*%(C_NS)s_SER_BYTES],
-    const %(c_ns)s_point_t pt,
+$(c_ns)_invert_elligator_uniform (
+    unsigned char recovered_hash[2*$(C_NS)_SER_BYTES],
+    const $(c_ns)_point_t pt,
     uint16_t which
 ) API_VIS NONNULL2 NOINLINE WARN_UNUSED;
 
 /**
  * @brief Overwrite scalar with zeros.
  */
-void %(c_ns)s_scalar_destroy (
-    %(c_ns)s_scalar_t scalar
+void $(c_ns)_scalar_destroy (
+    $(c_ns)_scalar_t scalar
 ) NONNULL1 API_VIS;
 
 /**
  * @brief Overwrite point with zeros.
  * @todo Use this internally.
  */
-void %(c_ns)s_point_destroy (
-    %(c_ns)s_point_t point
+void $(c_ns)_point_destroy (
+    $(c_ns)_point_t point
 ) NONNULL1 API_VIS;
 
 /**
  * @brief Overwrite precomputed table with zeros.
  */
-void %(c_ns)s_precomputed_destroy (
-    %(c_ns)s_precomputed_s *pre
+void $(c_ns)_precomputed_destroy (
+    $(c_ns)_precomputed_s *pre
 ) NONNULL1 API_VIS;
 
 #ifdef __cplusplus
 } /* extern "C" */
 #endif
-"""
-)
diff --git a/src/gen_headers/decaf_hxx.py b/src/per_curve/decaf.tmpl.hxx
similarity index 75%
rename from src/gen_headers/decaf_hxx.py
rename to src/per_curve/decaf.tmpl.hxx
index 94a91e3..c4cc01b 100644
--- a/src/gen_headers/decaf_hxx.py
+++ b/src/per_curve/decaf.tmpl.hxx
@@ -1,23 +1,18 @@
-from gen_file import gen_file
-
-decaf_hxx = gen_file(
-    public = True,
-    per = "curve",
-    name = "decaf/%(c_ns)s.hxx",
-    doc = """
-        A group of prime order p, C++ wrapper.
-
-        The Decaf library implements cryptographic operations on a an elliptic curve
-        group of prime order p. It accomplishes this by using a twisted Edwards
-        curve (isogenous to %(iso_to)s) and wiping out the cofactor.
-
-        The formulas are all complete and have no special cases, except that
-        %(c_ns)s_decode can fail because not every sequence of bytes is a valid group
-        element.
-
-        The formulas contain no data-dependent branches, timing or memory accesses,
-        except for %(c_ns)s_base_double_scalarmul_non_secret.
-    """, code = """
+/**
+ * A group of prime order p, C++ wrapper.
+ * 
+ * The Decaf library implements cryptographic operations on a an elliptic curve
+ * group of prime order p. It accomplishes this by using a twisted Edwards
+ * curve (isogenous to $(iso_to)) and wiping out the cofactor.
+ * 
+ * The formulas are all complete and have no special cases, except that
+ * $(c_ns)_decode can fail because not every sequence of bytes is a valid group
+ * element.
+ * 
+ * The formulas contain no data-dependent branches, timing or memory accesses,
+ * except for $(c_ns)_base_double_scalarmul_non_secret.
+ */
+
 /** This code uses posix_memalign. */
 #ifndef _XOPEN_SOURCE
 #define _XOPEN_SOURCE 600
@@ -25,7 +20,7 @@ decaf_hxx = gen_file(
 #include <stdlib.h>
 #include <string.h> /* for memcpy */
 
-#include <decaf.h>
+#include <decaf/decaf_$(gf_bits).h>
 #include <decaf/secure_buffer.hxx>
 #include <string>
 #include <sys/types.h>
@@ -42,18 +37,18 @@ decaf_hxx = gen_file(
 namespace decaf {
 
 /**
- * %(iso_to)s/Decaf instantiation of group.
+ * $(iso_to)/Decaf instantiation of group.
  */
-struct %(cxx_ns)s {
+struct $(cxx_ns) {
 
 /** The name of the curve */
-static inline const char *name() { return "%(name)s"; }
+static inline const char *name() { return "$(name)"; }
 
 /** The curve's cofactor (removed, but useful for testing) */
-static const int REMOVED_COFACTOR = %(cofactor)d;
+static const int REMOVED_COFACTOR = $(cofactor);
 
 /** Residue class of field modulus: p == this mod 2*(this-1) */
-static const int FIELD_MODULUS_TYPE = %(modulus_type)d;
+static const int FIELD_MODULUS_TYPE = $(modulus_type);
 
 /** @cond internal */
 class Point;
@@ -67,10 +62,10 @@ class Precomputed;
 class Scalar : public Serializable<Scalar> {
 public:
     /** wrapped C type */
-    typedef %(c_ns)s_scalar_t Wrapped;
+    typedef $(c_ns)_scalar_t Wrapped;
     
     /** Size of a serialized element */
-    static const size_t SER_BYTES = %(C_NS)s_SCALAR_BYTES;
+    static const size_t SER_BYTES = $(C_NS)_SCALAR_BYTES;
 
     /** access to the underlying scalar object */
     Wrapped s;
@@ -99,7 +94,7 @@ public:
     }
 
     /** Construct from decaf_scalar_t object. */
-    inline Scalar(const Wrapped &t = %(c_ns)s_scalar_zero) NOEXCEPT { %(c_ns)s_scalar_copy(s,t); }
+    inline Scalar(const Wrapped &t = $(c_ns)_scalar_zero) NOEXCEPT { $(c_ns)_scalar_copy(s,t); }
 
     /** Copy constructor. */
     inline Scalar(const Scalar &x) NOEXCEPT { *this = x; }
@@ -112,20 +107,20 @@ public:
 
     /** Serializable instance */
     inline void serialize_into(unsigned char *buffer) const NOEXCEPT {
-        %(c_ns)s_scalar_encode(buffer, s);
+        $(c_ns)_scalar_encode(buffer, s);
     }
 
     /** Assignment. */
-    inline Scalar& operator=(const Scalar &x) NOEXCEPT { %(c_ns)s_scalar_copy(s,x.s); return *this; }
+    inline Scalar& operator=(const Scalar &x) NOEXCEPT { $(c_ns)_scalar_copy(s,x.s); return *this; }
 
     /** Assign from unsigned 64-bit integer. */
-    inline Scalar& operator=(uint64_t w) NOEXCEPT { %(c_ns)s_scalar_set_unsigned(s,w); return *this; }
+    inline Scalar& operator=(uint64_t w) NOEXCEPT { $(c_ns)_scalar_set_unsigned(s,w); return *this; }
 
 
     /** Assign from signed int. */
     inline Scalar& operator=(int64_t w) NOEXCEPT {
         Scalar t(-(uint64_t)INT_MIN);
-        %(c_ns)s_scalar_set_unsigned(s,(uint64_t)w - (uint64_t)INT_MIN);
+        $(c_ns)_scalar_set_unsigned(s,(uint64_t)w - (uint64_t)INT_MIN);
         *this -= t;
         return *this;
     }
@@ -137,11 +132,11 @@ public:
     inline Scalar& operator=(int w) NOEXCEPT { return *this = (int64_t)w; }
 
     /** Destructor securely zeorizes the scalar. */
-    inline ~Scalar() NOEXCEPT { %(c_ns)s_scalar_destroy(s); }
+    inline ~Scalar() NOEXCEPT { $(c_ns)_scalar_destroy(s); }
 
     /** Assign from arbitrary-length little-endian byte sequence in a Block. */
     inline Scalar &operator=(const Block &bl) NOEXCEPT {
-        %(c_ns)s_scalar_decode_long(s,bl.data(),bl.size()); return *this;
+        $(c_ns)_scalar_decode_long(s,bl.data(),bl.size()); return *this;
     }
 
     /**
@@ -151,35 +146,35 @@ public:
     static inline decaf_error_t WARN_UNUSED decode (
         Scalar &sc, const FixedBlock<SER_BYTES> buffer
     ) NOEXCEPT {
-        return %(c_ns)s_scalar_decode(sc.s,buffer.data());
+        return $(c_ns)_scalar_decode(sc.s,buffer.data());
     }
 
     /** Add. */
-    inline Scalar operator+ (const Scalar &q) const NOEXCEPT { Scalar r((NOINIT())); %(c_ns)s_scalar_add(r.s,s,q.s); return r; }
+    inline Scalar operator+ (const Scalar &q) const NOEXCEPT { Scalar r((NOINIT())); $(c_ns)_scalar_add(r.s,s,q.s); return r; }
 
     /** Add to this. */
-    inline Scalar &operator+=(const Scalar &q) NOEXCEPT { %(c_ns)s_scalar_add(s,s,q.s); return *this; }
+    inline Scalar &operator+=(const Scalar &q) NOEXCEPT { $(c_ns)_scalar_add(s,s,q.s); return *this; }
 
     /** Subtract. */
-    inline Scalar operator- (const Scalar &q) const NOEXCEPT { Scalar r((NOINIT())); %(c_ns)s_scalar_sub(r.s,s,q.s); return r; }
+    inline Scalar operator- (const Scalar &q) const NOEXCEPT { Scalar r((NOINIT())); $(c_ns)_scalar_sub(r.s,s,q.s); return r; }
 
     /** Subtract from this. */
-    inline Scalar &operator-=(const Scalar &q) NOEXCEPT { %(c_ns)s_scalar_sub(s,s,q.s); return *this; }
+    inline Scalar &operator-=(const Scalar &q) NOEXCEPT { $(c_ns)_scalar_sub(s,s,q.s); return *this; }
 
     /** Multiply */
-    inline Scalar operator* (const Scalar &q) const NOEXCEPT { Scalar r((NOINIT())); %(c_ns)s_scalar_mul(r.s,s,q.s); return r; }
+    inline Scalar operator* (const Scalar &q) const NOEXCEPT { Scalar r((NOINIT())); $(c_ns)_scalar_mul(r.s,s,q.s); return r; }
 
     /** Multiply into this. */
-    inline Scalar &operator*=(const Scalar &q) NOEXCEPT { %(c_ns)s_scalar_mul(s,s,q.s); return *this; }
+    inline Scalar &operator*=(const Scalar &q) NOEXCEPT { $(c_ns)_scalar_mul(s,s,q.s); return *this; }
 
     /** Negate */
-    inline Scalar operator- () const NOEXCEPT { Scalar r((NOINIT())); %(c_ns)s_scalar_sub(r.s,%(c_ns)s_scalar_zero,s); return r; }
+    inline Scalar operator- () const NOEXCEPT { Scalar r((NOINIT())); $(c_ns)_scalar_sub(r.s,$(c_ns)_scalar_zero,s); return r; }
 
     /** Invert with Fermat's Little Theorem (slow!). If *this == 0,
      * throw CryptoException. */
     inline Scalar inverse() const throw(CryptoException) {
         Scalar r;
-        if (DECAF_SUCCESS != %(c_ns)s_scalar_invert(r.s,s)) {
+        if (DECAF_SUCCESS != $(c_ns)_scalar_invert(r.s,s)) {
             throw CryptoException();
         }
         return r;
@@ -189,7 +184,7 @@ public:
      * and return DECAF_FAILURE. */
     inline decaf_error_t WARN_UNUSED
     inverse_noexcept(Scalar &r) const NOEXCEPT {
-        return %(c_ns)s_scalar_invert(r.s,s);
+        return $(c_ns)_scalar_invert(r.s,s);
     }
 
     /** Divide by inverting q. If q == 0, return 0. */
@@ -202,7 +197,7 @@ public:
     inline bool operator!=(const Scalar &q) const NOEXCEPT { return !(*this == q); }
 
     /** Compare in constant time */
-    inline bool operator==(const Scalar &q) const NOEXCEPT { return !!%(c_ns)s_scalar_eq(s,q.s); }
+    inline bool operator==(const Scalar &q) const NOEXCEPT { return !!$(c_ns)_scalar_eq(s,q.s); }
 
     /** Scalarmul with scalar on left. */
     inline Point operator* (const Point &q) const NOEXCEPT { return q * (*this); }
@@ -224,10 +219,10 @@ public:
 class Point : public Serializable<Point> {
 public:
     /** wrapped C type */
-    typedef %(c_ns)s_point_t Wrapped;
+    typedef $(c_ns)_point_t Wrapped;
     
     /** Size of a serialized element */
-    static const size_t SER_BYTES = %(C_NS)s_SER_BYTES;
+    static const size_t SER_BYTES = $(C_NS)_SER_BYTES;
 
     /** Bytes required for hash */
     static const size_t HASH_BYTES = SER_BYTES;
@@ -244,16 +239,16 @@ public:
     /** @endcond */
 
     /** Constructor sets to identity by default. */
-    inline Point(const Wrapped &q = %(c_ns)s_point_identity) NOEXCEPT { %(c_ns)s_point_copy(p,q); }
+    inline Point(const Wrapped &q = $(c_ns)_point_identity) NOEXCEPT { $(c_ns)_point_copy(p,q); }
 
     /** Copy constructor. */
     inline Point(const Point &q) NOEXCEPT { *this = q; }
 
     /** Assignment. */
-    inline Point& operator=(const Point &q) NOEXCEPT { %(c_ns)s_point_copy(p,q.p); return *this; }
+    inline Point& operator=(const Point &q) NOEXCEPT { $(c_ns)_point_copy(p,q.p); return *this; }
 
     /** Destructor securely zeorizes the point. */
-    inline ~Point() NOEXCEPT { %(c_ns)s_point_destroy(p); }
+    inline ~Point() NOEXCEPT { $(c_ns)_point_destroy(p); }
 
     /** Construct from RNG */
     inline explicit Point(Rng &rng, bool uniform = true) NOEXCEPT {
@@ -291,7 +286,7 @@ public:
     static inline decaf_error_t WARN_UNUSED decode (
         Point &p, const FixedBlock<SER_BYTES> &buffer, decaf_bool_t allow_identity=DECAF_TRUE
     ) NOEXCEPT {
-        return %(c_ns)s_point_decode(p.p,buffer.data(),allow_identity);
+        return $(c_ns)_point_decode(p.p,buffer.data(),allow_identity);
     }
 
     /**
@@ -314,15 +309,15 @@ public:
         if (s.size() < HASH_BYTES) {
             SecureBuffer b(HASH_BYTES);
             memcpy(b.data(), s.data(), s.size());
-            %(c_ns)s_point_from_hash_nonuniform(p,b.data());
+            $(c_ns)_point_from_hash_nonuniform(p,b.data());
         } else if (s.size() == HASH_BYTES) {
-            %(c_ns)s_point_from_hash_nonuniform(p,s.data());
+            $(c_ns)_point_from_hash_nonuniform(p,s.data());
         } else if (s.size() < 2*HASH_BYTES) {
             SecureBuffer b(2*HASH_BYTES);
             memcpy(b.data(), s.data(), s.size());
-            %(c_ns)s_point_from_hash_uniform(p,b.data());
+            $(c_ns)_point_from_hash_uniform(p,b.data());
         } else {
-            %(c_ns)s_point_from_hash_uniform(p,s.data());
+            $(c_ns)_point_from_hash_uniform(p,s.data());
         }
     }
 
@@ -331,7 +326,7 @@ public:
      */
     inline operator SecureBuffer() const {
         SecureBuffer buffer(SER_BYTES);
-        %(c_ns)s_point_encode(buffer.data(), p);
+        $(c_ns)_point_encode(buffer.data(), p);
         return buffer;
     }
 
@@ -340,41 +335,41 @@ public:
 
     /** Serializable instance */
     inline void serialize_into(unsigned char *buffer) const NOEXCEPT {
-        %(c_ns)s_point_encode(buffer, p);
+        $(c_ns)_point_encode(buffer, p);
     }
 
     /** Point add. */
-    inline Point operator+ (const Point &q) const NOEXCEPT { Point r((NOINIT())); %(c_ns)s_point_add(r.p,p,q.p); return r; }
+    inline Point operator+ (const Point &q) const NOEXCEPT { Point r((NOINIT())); $(c_ns)_point_add(r.p,p,q.p); return r; }
 
     /** Point add. */
-    inline Point &operator+=(const Point &q) NOEXCEPT { %(c_ns)s_point_add(p,p,q.p); return *this; }
+    inline Point &operator+=(const Point &q) NOEXCEPT { $(c_ns)_point_add(p,p,q.p); return *this; }
 
     /** Point subtract. */
-    inline Point operator- (const Point &q) const NOEXCEPT { Point r((NOINIT())); %(c_ns)s_point_sub(r.p,p,q.p); return r; }
+    inline Point operator- (const Point &q) const NOEXCEPT { Point r((NOINIT())); $(c_ns)_point_sub(r.p,p,q.p); return r; }
 
     /** Point subtract. */
-    inline Point &operator-=(const Point &q) NOEXCEPT { %(c_ns)s_point_sub(p,p,q.p); return *this; }
+    inline Point &operator-=(const Point &q) NOEXCEPT { $(c_ns)_point_sub(p,p,q.p); return *this; }
 
     /** Point negate. */
-    inline Point operator- () const NOEXCEPT { Point r((NOINIT())); %(c_ns)s_point_negate(r.p,p); return r; }
+    inline Point operator- () const NOEXCEPT { Point r((NOINIT())); $(c_ns)_point_negate(r.p,p); return r; }
 
     /** Double the point out of place. */
-    inline Point times_two () const NOEXCEPT { Point r((NOINIT())); %(c_ns)s_point_double(r.p,p); return r; }
+    inline Point times_two () const NOEXCEPT { Point r((NOINIT())); $(c_ns)_point_double(r.p,p); return r; }
 
     /** Double the point in place. */
-    inline Point &double_in_place() NOEXCEPT { %(c_ns)s_point_double(p,p); return *this; }
+    inline Point &double_in_place() NOEXCEPT { $(c_ns)_point_double(p,p); return *this; }
 
     /** Constant-time compare. */
-    inline bool operator!=(const Point &q) const NOEXCEPT { return ! %(c_ns)s_point_eq(p,q.p); }
+    inline bool operator!=(const Point &q) const NOEXCEPT { return ! $(c_ns)_point_eq(p,q.p); }
 
     /** Constant-time compare. */
-    inline bool operator==(const Point &q) const NOEXCEPT { return !!%(c_ns)s_point_eq(p,q.p); }
+    inline bool operator==(const Point &q) const NOEXCEPT { return !!$(c_ns)_point_eq(p,q.p); }
 
     /** Scalar multiply. */
-    inline Point operator* (const Scalar &s) const NOEXCEPT { Point r((NOINIT())); %(c_ns)s_point_scalarmul(r.p,p,s.s); return r; }
+    inline Point operator* (const Scalar &s) const NOEXCEPT { Point r((NOINIT())); $(c_ns)_point_scalarmul(r.p,p,s.s); return r; }
 
     /** Scalar multiply in place. */
-    inline Point &operator*=(const Scalar &s) NOEXCEPT { %(c_ns)s_point_scalarmul(p,p,s.s); return *this; }
+    inline Point &operator*=(const Scalar &s) NOEXCEPT { $(c_ns)_point_scalarmul(p,p,s.s); return *this; }
 
     /** Multiply by s.inverse(). If s=0, maps to the identity. */
     inline Point operator/ (const Scalar &s) const throw(CryptoException) { return (*this) * s.inverse(); }
@@ -383,20 +378,20 @@ public:
     inline Point &operator/=(const Scalar &s) throw(CryptoException) { return (*this) *= s.inverse(); }
 
     /** Validate / sanity check */
-    inline bool validate() const NOEXCEPT { return %(c_ns)s_point_valid(p); }
+    inline bool validate() const NOEXCEPT { return $(c_ns)_point_valid(p); }
 
     /** Double-scalar multiply, equivalent to q*qs + r*rs but faster. */
     static inline Point double_scalarmul (
         const Point &q, const Scalar &qs, const Point &r, const Scalar &rs
     ) NOEXCEPT {
-        Point p((NOINIT())); %(c_ns)s_point_double_scalarmul(p.p,q.p,qs.s,r.p,rs.s); return p;
+        Point p((NOINIT())); $(c_ns)_point_double_scalarmul(p.p,q.p,qs.s,r.p,rs.s); return p;
     }
 
     /** Dual-scalar multiply, equivalent to this*r1, this*r2 but faster. */
     inline void dual_scalarmul (
         Point &q1, Point &q2, const Scalar &r1, const Scalar &r2
     ) const NOEXCEPT {
-        %(c_ns)s_point_dual_scalarmul(q1.p,q2.p,p,r1.s,r2.s);
+        $(c_ns)_point_dual_scalarmul(q1.p,q2.p,p,r1.s,r2.s);
     }
 
     /**
@@ -415,20 +410,20 @@ public:
      * it doesn't).
      */
     inline Point non_secret_combo_with_base(const Scalar &s, const Scalar &s_base) NOEXCEPT {
-        Point r((NOINIT())); %(c_ns)s_base_double_scalarmul_non_secret(r.p,s_base.s,p,s.s); return r;
+        Point r((NOINIT())); $(c_ns)_base_double_scalarmul_non_secret(r.p,s_base.s,p,s.s); return r;
     }
 
     /** Return a point equal to *this, whose internal data is rotated by a torsion element. */
     inline Point debugging_torque() const NOEXCEPT {
         Point q;
-        %(c_ns)s_point_debugging_torque(q.p,p);
+        $(c_ns)_point_debugging_torque(q.p,p);
         return q;
     }
 
     /** Return a point equal to *this, whose internal data has a modified representation. */
     inline Point debugging_pscale(const FixedBlock<SER_BYTES> factor) const NOEXCEPT {
         Point q;
-        %(c_ns)s_point_debugging_pscale(q.p,p,factor.data());
+        $(c_ns)_point_debugging_pscale(q.p,p,factor.data());
         return q;
     }
 
@@ -450,9 +445,9 @@ public:
         memcpy(buf2,buf.data(),(buf.size() > 2*HASH_BYTES) ? 2*HASH_BYTES : buf.size());
         decaf_bool_t ret;
         if (buf.size() > HASH_BYTES) {
-            ret = decaf_successful(%(c_ns)s_invert_elligator_uniform(buf2, p, hint));
+            ret = decaf_successful($(c_ns)_invert_elligator_uniform(buf2, p, hint));
         } else {
-            ret = decaf_successful(%(c_ns)s_invert_elligator_nonuniform(buf2, p, hint));
+            ret = decaf_successful($(c_ns)_invert_elligator_nonuniform(buf2, p, hint));
         }
         if (buf.size() < HASH_BYTES) {
             ret &= decaf_memeq(&buf2[buf.size()], &buf2[HASH_BYTES], HASH_BYTES - buf.size());
@@ -477,10 +472,10 @@ public:
     }
 
     /** Return the base point */
-    static inline const Point base() NOEXCEPT { return Point(%(c_ns)s_point_base); }
+    static inline const Point base() NOEXCEPT { return Point($(c_ns)_point_base); }
 
     /** Return the identity point */
-    static inline const Point identity() NOEXCEPT { return Point(%(c_ns)s_point_identity); }
+    static inline const Point identity() NOEXCEPT { return Point($(c_ns)_point_identity); }
 };
 
 /**
@@ -491,7 +486,7 @@ public:
  */
 
 /** @cond internal */
-typedef %(c_ns)s_precomputed_s Precomputed_U;
+typedef $(c_ns)_precomputed_s Precomputed_U;
 /** @endcond */
 class Precomputed
     /** @cond internal */
@@ -543,7 +538,7 @@ public:
      */
     inline Precomputed &operator=(const Point &it) throw(std::bad_alloc) {
         alloc();
-        %(c_ns)s_precompute(ours.mine,it.p);
+        $(c_ns)_precompute(ours.mine,it.p);
         return *this;
     }
 
@@ -560,7 +555,7 @@ public:
         : OwnedOrUnowned<Precomputed,Precomputed_U>() { *this = it; }
 
     /** Fixed base scalarmul. */
-    inline Point operator* (const Scalar &s) const NOEXCEPT { Point r; %(c_ns)s_precomputed_scalarmul(r.p,get(),s.s); return r; }
+    inline Point operator* (const Scalar &s) const NOEXCEPT { Point r; $(c_ns)_precomputed_scalarmul(r.p,get(),s.s); return r; }
 
     /** Multiply by s.inverse(). If s=0, maps to the identity. */
     inline Point operator/ (const Scalar &s) const throw(CryptoException) { return (*this) * s.inverse(); }
@@ -571,23 +566,23 @@ public:
 public:
     /** @cond internal */
     friend class OwnedOrUnowned<Precomputed,Precomputed_U>;
-    static inline size_t size() NOEXCEPT { return %(c_ns)s_sizeof_precomputed_s; }
-    static inline size_t alignment() NOEXCEPT { return %(c_ns)s_alignof_precomputed_s; }
-    static inline const Precomputed_U * defaultValue() NOEXCEPT { return %(c_ns)s_precomputed_base; }
+    static inline size_t size() NOEXCEPT { return $(c_ns)_sizeof_precomputed_s; }
+    static inline size_t alignment() NOEXCEPT { return $(c_ns)_alignof_precomputed_s; }
+    static inline const Precomputed_U * defaultValue() NOEXCEPT { return $(c_ns)_precomputed_base; }
     /** @endcond */
 };
 
 struct DhLadder {
 public:
-    /** Bytes in an X%(gf_shortname)s public key. */
-    static const size_t PUBLIC_BYTES = X%(gf_shortname)s_PUBLIC_BYTES;
+    /** Bytes in an X$(gf_shortname) public key. */
+    static const size_t PUBLIC_BYTES = X$(gf_shortname)_PUBLIC_BYTES;
 
-    /** Bytes in an X%(gf_shortname)s private key. */
-    static const size_t PRIVATE_BYTES = X%(gf_shortname)s_PRIVATE_BYTES;
+    /** Bytes in an X$(gf_shortname) private key. */
+    static const size_t PRIVATE_BYTES = X$(gf_shortname)_PRIVATE_BYTES;
 
     /** Base point for a scalar multiplication. */
     static const FixedBlock<PUBLIC_BYTES> base_point() NOEXCEPT {
-        return FixedBlock<PUBLIC_BYTES>(%(c_ns)s_x_base_point);
+        return FixedBlock<PUBLIC_BYTES>($(c_ns)_x_base_point);
     }
 
     /** Generate and return a shared secret with public key.  */
@@ -596,7 +591,7 @@ public:
         const FixedBlock<PRIVATE_BYTES> &scalar
     ) throw(std::bad_alloc,CryptoException) {
         SecureBuffer out(PUBLIC_BYTES);
-        if (DECAF_SUCCESS != %(c_ns)s_x_direct_scalarmul(out.data(), pk.data(), scalar.data())) {
+        if (DECAF_SUCCESS != $(c_ns)_x_direct_scalarmul(out.data(), pk.data(), scalar.data())) {
             throw CryptoException();
         }
         return out;
@@ -609,7 +604,7 @@ public:
         const FixedBlock<PUBLIC_BYTES> &pk,
         const FixedBlock<PRIVATE_BYTES> &scalar
     ) NOEXCEPT {
-       return %(c_ns)s_x_direct_scalarmul(out.data(), pk.data(), scalar.data());
+       return $(c_ns)_x_direct_scalarmul(out.data(), pk.data(), scalar.data());
     }
 
     /** Generate and return a public key; equivalent to shared_secret(base_point(),scalar)
@@ -619,7 +614,7 @@ public:
         const FixedBlock<PRIVATE_BYTES> &scalar
     ) throw(std::bad_alloc) {
         SecureBuffer out(PUBLIC_BYTES);
-        %(c_ns)s_x_base_scalarmul(out.data(), scalar.data());
+        $(c_ns)_x_base_scalarmul(out.data(), scalar.data());
         return out;
     }
 
@@ -631,21 +626,21 @@ public:
         FixedBuffer<PUBLIC_BYTES> &out,
         const FixedBlock<PRIVATE_BYTES> &scalar
     ) NOEXCEPT {
-        %(c_ns)s_x_base_scalarmul(out.data(), scalar.data());
+        $(c_ns)_x_base_scalarmul(out.data(), scalar.data());
     }
 };
 
-}; /* struct %(cxx_ns)s */
+}; /* struct $(cxx_ns) */
 
 /** @cond internal */
-inline SecureBuffer %(cxx_ns)s::Scalar::direct_scalarmul (
+inline SecureBuffer $(cxx_ns)::Scalar::direct_scalarmul (
     const Block &in,
     decaf_bool_t allow_identity,
     decaf_bool_t short_circuit
 ) const throw(CryptoException) {
-    SecureBuffer out(%(cxx_ns)s::Point::SER_BYTES);
+    SecureBuffer out($(cxx_ns)::Point::SER_BYTES);
     if (DECAF_SUCCESS !=
-        %(c_ns)s_direct_scalarmul(out.data(), in.data(), s, allow_identity, short_circuit)
+        $(c_ns)_direct_scalarmul(out.data(), in.data(), s, allow_identity, short_circuit)
     ) {
         throw CryptoException();
     }
@@ -655,6 +650,3 @@ inline SecureBuffer %(cxx_ns)s::Scalar::direct_scalarmul (
 
 #undef NOEXCEPT
 } /* namespace decaf */
-"""
-)
-
diff --git a/src/public_include/decaf.tmpl.h b/src/public_include/decaf.tmpl.h
new file mode 100644
index 0000000..c4c1fe2
--- /dev/null
+++ b/src/public_include/decaf.tmpl.h
@@ -0,0 +1,18 @@
+/**
+ * Master header for Decaf library.
+ * 
+ * The Decaf library implements cryptographic operations on a elliptic curve
+ * groups of prime order p.  It accomplishes this by using a twisted Edwards
+ * curve (isogenous to Ed448-Goldilocks or Ed25519) and wiping out the cofactor.
+ * 
+ * The formulas are all complete and have no special cases.  However, some
+ * functions can fail.  For example, decoding functions can fail because not
+ * every string is the encoding of a valid group element.
+ * 
+ * The formulas contain no data-dependent branches, timing or memory accesses,
+ * except for decaf_XXX_base_double_scalarmul_non_secret.
+ */
+
+$("\n".join([
+    "#include <decaf/decaf_%s.h>" % g for g in sorted([c["bits"] for _,c in curve.iteritems()])
+]))
diff --git a/src/public_include/decaf.tmpl.hxx b/src/public_include/decaf.tmpl.hxx
new file mode 100644
index 0000000..2da5246
--- /dev/null
+++ b/src/public_include/decaf.tmpl.hxx
@@ -0,0 +1,16 @@
+/** Master header for Decaf library, C++ version. */
+
+$("\n".join([
+    "#include <decaf/decaf_%s.hxx>" % g for g in sorted([c["bits"] for _,c in curve.iteritems()])
+]))
+
+namespace decaf {
+    template <template<typename Group> class Run>
+    void run_for_all_curves() {
+$("\n".join([
+"        Run<%s>::run();" % cd["cxx_ns"]
+for cd in sorted(curve.values(), key=lambda x:x["c_ns"])
+])
+)
+    }
+}
diff --git a/src/public_include/decaf/crypto.tmpl.h b/src/public_include/decaf/crypto.tmpl.h
new file mode 100644
index 0000000..dec7617
--- /dev/null
+++ b/src/public_include/decaf/crypto.tmpl.h
@@ -0,0 +1,10 @@
+/**
+ * Example Decaf crypto routines, metaheader.
+ * @warning These are merely examples, though they ought to be secure.  But real
+ * protocols will decide differently on magic numbers, formats, which items to
+ * hash, etc.
+ */
+
+$("\n".join([
+    "#include <decaf/crypto_%s.h>" % g for g in sorted([c["bits"] for _,c in curve.iteritems()])
+]))
diff --git a/src/public_include/decaf/crypto.tmpl.hxx b/src/public_include/decaf/crypto.tmpl.hxx
new file mode 100644
index 0000000..be98c6f
--- /dev/null
+++ b/src/public_include/decaf/crypto.tmpl.hxx
@@ -0,0 +1,10 @@
+/**
+ * Example Decaf crypto routines, C++ metaheader.
+ * @warning These are merely examples, though they ought to be secure.  But real
+ * protocols will decide differently on magic numbers, formats, which items to
+ * hash, etc.
+ */
+
+$("\n".join([
+    "#include <decaf/crypto_%s.hxx>" % g for g in sorted([c["bits"] for _,c in curve.iteritems()])
+]))