diff --git a/backend/scripts/build_reference.py b/backend/scripts/build_reference.py
index da95850..fe77fbe 100644
--- a/backend/scripts/build_reference.py
+++ b/backend/scripts/build_reference.py
@@ -20,6 +20,105 @@
 _BASE_MODEL_DOC = inspect.getdoc(BaseModel) or ""
 
 
+# ---------------------------------------------------------------------------
+# Canonical reform recipes
+# ---------------------------------------------------------------------------
+# Copy-pasteable Python the agent can adapt verbatim inside `run_python`.
+# Critical lesson from production: `Parameters(income_tax={"uk_brackets": [...]})`
+# REPLACES the full schedule — passing a single-bracket list zeroes out the
+# higher and additional rates and tanks revenue. To raise one rate you must
+# pass the FULL bracket list with only that rate changed (recipe 1 below).
+#
+# Coverage: keys verified against backend/tests/test_agent_tools.py are marked
+# VERIFIED. The remaining recipes use the conventional PolicyEngine UK naming
+# from `_build_compiled_policy` (backend/agent_tools.py) — they should validate
+# against `pe.Parameters.model_json_schema()` which is dumped further down in
+# this reference. If a field name below is rejected, look it up in the JSON
+# schema section and use the closest match.
+REFORM_RECIPES = '''## Reform recipes
+
+Canonical, copy-pasteable reforms. Each recipe constructs a *partial* override
+on top of the current-law baseline — do not invent your own bracket lists from
+scratch unless you really want to replace an entire schedule.
+
+If a field name below is rejected by `Parameters(...)`, search the Parameters
+JSON schema section of this reference for the closest match — programme keys
+(`income_tax`, `national_insurance`, `child_benefit`, ...) and field names are
+authoritative there.
+
+### 1. Basic rate +1pp (20% -> 21%) — VERIFIED
+
+`uk_brackets` replaces the full bracket list, so always pass every band with
+just the one rate changed. This is the recipe the agent was failing on in
+production.
+
+```python
+reform = Parameters(income_tax={"uk_brackets": [
+    {"rate": 0.21, "threshold": 0.0},        # basic (was 0.20)
+    {"rate": 0.40, "threshold": 37700.0},    # higher (unchanged)
+    {"rate": 0.45, "threshold": 125140.0},   # additional (unchanged)
+]})
+sim = Simulation(year=2025, dataset="efrs")
+result = sim.run(policy=reform).budgetary_impact.model_dump()
+```
+
+### 2. Personal allowance £12,570 -> £15,000 — VERIFIED
+
+```python
+reform = Parameters(income_tax={"personal_allowance": 15000})
+sim = Simulation(year=2025, dataset="efrs")
+result = sim.run(policy=reform).budgetary_impact.model_dump()
+```
+
+### 3. NI primary threshold +£1,000
+
+```python
+# Verify field name against the Parameters JSON schema for `national_insurance`
+# if this is rejected; primary-threshold-style names vary by release.
+reform = Parameters(national_insurance={"primary_threshold": 13570})
+sim = Simulation(year=2025, dataset="efrs")
+result = sim.run(policy=reform).budgetary_impact.model_dump()
+```
+
+### 4. Child benefit +10% uprating
+
+```python
+# Verify the exact rate field name against the `child_benefit` schema;
+# common names are `amount_for_first_child` / `amount_for_additional_child`.
+reform = Parameters(child_benefit={
+    "amount_for_first_child": 26.95,        # 24.50 * 1.10 (weekly £)
+    "amount_for_additional_child": 17.85,   # 16.95 * 1.10 (weekly £)
+})
+sim = Simulation(year=2025, dataset="efrs")
+result = sim.run(policy=reform).budgetary_impact.model_dump()
+```
+
+### 5. Marriage allowance — turn off
+
+```python
+# Marriage allowance lives inside `income_tax` in the compiled engine.
+# Verify the exact field name (e.g. `marriage_allowance_fraction` or
+# `marriage_allowance`) against the Parameters JSON schema before relying on
+# this for production analysis.
+reform = Parameters(income_tax={"marriage_allowance": 0.0})
+sim = Simulation(year=2025, dataset="efrs")
+result = sim.run(policy=reform).budgetary_impact.model_dump()
+```
+
+### Patterns to apply across all recipes
+
+- `Parameters(programme={...})` is a *partial* override on the named programme:
+  fields you omit keep their current-law values. The one exception is list-
+  valued fields like `uk_brackets`, which are replaced wholesale — always
+  spell out the full list with only the rates/thresholds you intended to move.
+- For distributional output, use `sim.run(policy=reform).decile_impacts` and
+  `.winners_losers`; for revenue, `.budgetary_impact`.
+- `dataset="efrs"` is the Enhanced FRS — usually the default for distributional
+  work. Check `capabilities()` for the live list.
+'''
+
+
+
 def _own_doc(obj) -> str:
     """Return the object's own docstring, ignoring docs inherited from bases.
 
@@ -116,6 +215,10 @@ def render() -> str:
             lines.append(f"```python\n{name} = {value_repr}\n```")
             lines.append("")
 
+    # --- Reform recipes (canonical patterns) ---
+    lines.append(REFORM_RECIPES)
+    lines.append("")
+
     # --- Parameters schema ---
     lines.append("## Parameters JSON schema")
     lines.append("")