Create Implementation Plan

Planning only — no code changes, no execution. You produce one document: an implementation plan that is the contract for a later execution session.

This is the "how", not the "what". The "what" was settled in an earlier refinement step and lives in the requirements document you're given; do not redefine scope. But you must flag any gap, ambiguity, or inconsistency you find in the requirements — surface it, never paper over it.

Golden rule

Never guess — ask. With one limit: anything resolvable by reading the codebase, resolve by reading the codebase; only questions the code cannot settle go to the user.

Steps

Your task is to produce the implementation plan. The steps below build toward it; grilling the user — interviewing relentlessly to resolve anything the code can't settle — is woven through the design, not a separate phase that runs before planning starts.

1. Read the requirements document the user references (e.g. a *.REQUIREMENTS.md). If the path is ambiguous, ask. Its Open questions are forks to grill in step 3, each with its tentative answer as the recommendation; none stays open in the plan. Its Overrides carry over into part 2's Overrides.
2. Verify against the actual codebase. Open the files the requirements cite and confirm the prior-art references still hold; note anything that has shifted since the requirements were written. The requirements may also carry verified codebase facts — use any that are there to save re-discovery work, aware the code may have moved meanwhile (a pinned commit makes the check cheap: git diff <commit>..HEAD -- <cited paths>). If the requirements gate implementation on missing data or upstream work, verify that gate independently — stale gating claims are a common failure mode and easily inflate into a plan's first step when the data is in fact already addressable.

3. Work out the approach. This is the core of the task: design the "how" — what existing code to reuse, what to introduce, where each change goes, and the order of operations that avoids broken intermediate states (data model before its consumers, code before its tests, a bug's regression test aside: see the Plan file's Steps). Track dependencies between steps. Working this out surfaces the decision tree: the forks where more than one reasonable approach exists. Typical forks to design through (and grill on when the code can't settle them): whether to refactor existing code to reuse it or build anew; which API or interface to call; which unit tests to add; code style, file names, and folder structure.

   Docs generally outrank prior art. A mirrored analogue doesn't override a documented rule, and a rule's rationale doesn't license exceptions to it — existing code may predate the rule. When docs and prior art conflict and you're unsure how to resolve it, ask the user; a deviation the user approves is recorded in the plan's Overrides section, so the execution session knows it's deliberate.

   Prior art models shape, not content. Its settings fit its own runtime, tooling and purpose and may predate the current ones, so before a step reuses a file, check each setting against the new component's target and prescribe only what fits, citing the file for shape; "copy X verbatim" is a decision settled line by line, never by the file's existence. A technical claim a step rests on ("X needs exactly Y") is verified against a doc, a compile or a test, never read off what the prior art happens to do.

   Deletions tend to cascade. When a step deletes code, look for what that code was the last user of (a module, helper, template, constant, package) by searching for each one's other users, and consider naming it for deletion too, repeating for every piece newly orphaned; one the search cannot clear (an export, a name resolved at runtime) is a fork to grill. Each test deleted with its code gets a verdict by behaviour: the part that checks something the surviving code still does moves to a test the step names in its Where, and only the rest is deleted. A step that moves or renames code searches docs, agent instructions, CI workflows, build scripts and config for the old name in every form it is written (case and separator variants) and lists every hit in its Where.

   Walk every point a run can stop when steps change live state in stages (a deploy, a migration, a swap): each step, an abort, a timeout, a lost worker, an automatic re-run or restart. At each, settle what the target is left in and what the log, notifications and any manual-fix advice say there. Hazards recorded in related plans' decisions logs are prior art.

   Parts released separately each meet the other's live version (a backend and its frontend, a service and its callers, a schema and its code): prefer a shape it tolerates — additions optional, defaulting to the pre-change behaviour; removals and renames only once no live version still uses them. None reasonable → a fork to grill; the plan's Conventions then state the release order.

   Grill the user along the way. Whenever the approach hits a fork you can't settle from the code, stop and resolve it with the user before continuing — don't guess, and don't defer the decision into the plan. Interview relentlessly until you reach shared understanding, walking each branch of the decision tree and resolving dependencies between decisions:

   • One question at a time, each with your recommended answer, worded so that "yes" accepts it, via explain-in-simple-language when available.
   • Facts are yours to find (codebase, docs, tool output): look them up, never ask. Decisions are the user's: put each one to them and wait.
   • Cover every gap, ambiguity, or inconsistency surfaced while reading, verifying, and designing.
4. Write the plan (below) — once the approach is settled, no open question remains, every step serves an AC or a convention, and identifiers and signatures match across steps.

Plan file

The output must be self-contained and ready for a fresh session that reads only the plan file and starts implementing — it should not need to read the requirements document or the ticket. This is the single most important constraint. Anything the execution session needs (acceptance criteria, prior-art citations, file paths, naming/string conventions, type signatures, override notes) must appear in the plan itself. Repetition from the requirements is intentional: the plan is the contract, not a diff against the requirements.

Create the file in the same directory as the requirements document, named by replacing .REQUIREMENTS with .PLAN (e.g. FOO.REQUIREMENTS.md → FOO.PLAN.md). If the input doesn't follow that convention, append .PLAN before .md. <id>-<slug> is the plan filename without .PLAN.md (123-report-approval); the id is absent when the task has no ticket.

Structure — five parts:

1. Summary — 1–3 sentences: what this plan implements and the overall shape of the change (which areas are touched).
2. Conventions and overrides — two distinct kinds of cross-cutting context:
   • Conventions: open with the commit the citations were verified at (short hash, dirty-tree note); then rules that apply across all steps — naming patterns, user-facing string rules, the testing approach (what to test, following existing test conventions in the codebase). Copy from the requirements verbatim where applicable.
   • Overrides: one-time corrections where the requirements deliberately diverge from a default or prior assumption. Each entry: what was assumed, what the actual requirement is, and why. The execution session applies each only where it fits.

3. Steps — ordered implementation steps. Each is concrete enough to act on but doesn't dictate every line, and is one change a fresh session finishes with context to spare; a bigger change is several steps. Each step has:

   • What — a short imperative.
   • Where — concrete file paths and identifiers.
   • How — the approach in prose. Reuse decisions go here, with path:line-range citations for prior art.
   • Snippet — only when prose is genuinely more ambiguous than code (a non-obvious type signature, a tricky nested structure, an unfamiliar call shape). Default to omitting; keep under ~10 lines.
   • Depends on — earlier step numbers this one requires, if any.

   Each step is self-contained — readable without scrolling back. No "as in step 2" without restating what step 2 did. Step 1 must never be discovery work ("find the field names", "explore where this lives") — discovery already happened in the steps above. A small deterministic sanity check of a known fact is fine; open-ended investigation is not.

   When the requirements carry a Reproduction part, step 1 is a regression test or a scripted reproduction asserting the expected behaviour that part records, whether or not the bug was reproduced, so it is not discovery; the fix follows in later steps. It is the smallest test that shows the bug (unit before integration before e2e), a larger one only when the bug lies in how parts connect. Its How says (the executor may know no such rule) that the test must be seen failing on the bug before the fix is written (its assertion, or the error the bug raises; a broken import or fixture is no failure seen), the failure seen is reported in one line, and a test that cannot be made to fail that way stops the run before any fix, with the reason reported. The step writing the fix says to rerun that test afterwards and see it pass.

4. Acceptance criteria — flat - [ ] AC-n checklist the implementation must satisfy, copied from the requirements' AC section. If unnumbered, number them (AC-1, …). Each references the step number(s) that satisfy it (e.g. "AC-3 → steps 4, 7"). Tag any AC that needs manual verification only (manual — leave unticked, report as pending). The execution session uses this as its done-check and ticks the rest. A "none left" AC (a removal, a doc sweep) carries its exact search — each removed name as a whole word, in every form it is written (case and separator variants) — with an allow-list of legitimate hits and the hits outside that list it finds on the default branch today, as a regression test is seen failing first (Steps). When the requirements give no search or no hits, the planner runs the search on the default branch and records them here; one that finds none is widened to the forms the name is actually written in, and still none means the AC records that nothing is left to remove.
5. Decisions log — a standing instruction copied into the plan for the execution session: "When a settled decision deviates from or extends this plan, append one line — the decision and its why — to <id>-<slug>.DECISIONS.md beside this plan, the moment it's settled." Without it, the why behind mid-implementation pivots is lost to later handover and review.

The plan must capture all the thinking — every decision needed to implement the feature, so the execution session re-derives nothing. It may omit only mechanical, locally-reversible detail with no cross-cutting consequence (a variable name, the exact wording of a log line). Anything that spans files, depends on knowing the codebase, or is costly to get wrong (which utility to reuse, what order to change things in, what new types to introduce) belongs in the plan. When unsure whether something is a decision or a mechanical detail, put it in.

Boundaries

• Do not modify any source files. The only file you write is the plan.
• Do not start implementing.
• When preparing the plan from a *.REQUIREMENTS.md file, you normally shouldn't need to go back to the *.TICKET.md — close any gaps with the user instead. Never assume a requirement; ask.
• Write the plan to disk only after every open question is resolved.

When done, state that the plan file is ready, by its project-relative path (relative to the working directory, never absolute), and offer these ways forward, each as its own fenced block under the label quoted below, holding a single copy-pasteable launch command — session name and prompt combined, so one paste starts a fresh session with a clean context. Use the agent tool's launch syntax (claude below is only the example) and name the session after <slug>, the plan filename's slug (without id prefix or extension), so the phase is recognizable in the session list:

1. "Execute the plan:" — reads only the plan file. <validation> is the project's own checks: the commands its governing docs name, else the toolchain's (lint, tests, build — only those that exist); nothing runnable → drop the clause:

   claude --name execute-plan-report-approval "Execute the plan .agents/plans/123-report-approval/123-report-approval.PLAN.md, then run <validation>; report, never commit"

2. "or split it into tasks first, for small reviewable increments:" — when split-plan-tasks is available:

   claude --name split-plan-tasks-report-approval "/split-plan-tasks .agents/plans/123-report-approval/123-report-approval.PLAN.md"

3. "or harden it first, so a fresh reviewer challenges it against the ticket and the code before anything is built:" — when harden-artifact is available:

   claude --name harden-plan-report-approval "/harden-artifact .agents/plans/123-report-approval/123-report-approval.PLAN.md"

Then offer the alternative: one line offering to clear the current session (/clear below is only the example; use the agent tool's clear command) and run the next phase here, then the same labelled blocks with the bare prompts.