Codex Final Execution Prompt

Codex Final Execution Prompt

Purpose

Write strict final execution prompts for Codex, Claude, Copilot, or other coding agents. A final execution prompt has one singular, bounded goal; every instruction in it is testable; and a stop condition exists for every ambiguous or blocked state so the agent never silently guesses. The output is a self-contained prompt — exact context, exact changes, hard constraints, verification commands, acceptance checklist, output contract, and stop conditions — ready to paste into a coding agent.

When to use

• A well-understood task must be delegated to a coding agent and the instruction set must be unambiguous enough to execute without back-and-forth.
• A previous agent run went off-scope, touched forbidden files, or failed to verify its work — write a tighter prompt with explicit hard constraints and stop conditions before re-running.
• A plan exists (from 04-implementation-plan) and now needs translating into a strict agent-executable format with exact file/change pairs and verification commands.
• A task involves risk (migration, rename, dependency change) and the prompt needs baked-in forbidden-action lists and rollback conditions to prevent silent damage.

When not to use

• The task is unrelated to product, docs, and research work.
• The work would require production deploys, destructive data actions, or secret disclosure.
• A narrower skill already covers the need: 63-md-spec-generator for human-facing specs and handoff packs, 04-implementation-plan for the planning phase that precedes prompt-writing.

Procedure

1. Lock the context. Target repo/paths, package manager, framework, runtime, and the exact, singular goal. One job, clearly bounded — not a basket of loosely related asks.
2. Enumerate constraints up front. What must not change, files/areas off-limits, style and conventions to follow, and forbidden actions (deploys, destructive git/DB, new deps, scope creep).
3. Specify the work concretely. Exact files to touch, the change in each, and the order if it matters.
4. Specify verification. The precise commands the agent must run and the expected green result for each.
5. Define acceptance criteria as a checklist the agent must satisfy before declaring done.
6. Define the output contract: what the agent must report back (diff summary, commands run, results, residual risks).
7. Strip ambiguity. No "etc.", no "improve as you see fit" — every instruction is testable.

Concrete checks

• The goal is singular and testable, not a bundle of related tasks.
• Every "do this" item names a concrete file and change, not a vibe ("clean up the code").
• Verification commands are exact and each has an expected result (tsc --noEmit → 0 errors).
• Forbidden and destructive actions are listed explicitly (no deploys, no git push --force, no DROP, no new dependencies).
• A stop condition exists for every blocked or ambiguous state, so the agent reports instead of guessing.
• The output contract is specified so the result is reviewable (files changed, commands run, residual risks).
• Off-limits paths are named, not implied.
• Secrets are referenced by env-var name only; no values are embedded.

Commands or Templates

# Task: <single, specific goal>

## Context
- Repo/paths: <...>   PM: <npm/pnpm/yarn/...>   Framework/runtime: <...>
- Relevant files: <path:line ...>

## Do exactly this
1. <change> in <file>
2. <change> in <file>

## Constraints (hard)
- Do NOT: deploy, run prod migrations, reset/seed data, force push,
  add dependencies, or touch <off-limits paths>.
- Preserve: <public API / behavior / existing patterns>.
- Match existing style and conventions.

## Verify (must pass before done)
- `<typecheck cmd>`  → clean (0 errors)
- `<test cmd>`       → green
- `<build cmd>`      → succeeds

## Acceptance criteria
- [ ] <observable outcome>
- [ ] <observable outcome>

## Report back
- Files changed (path → what), commands run + results,
  residual risks, anything skipped + why.

## Stop conditions
- If a constraint blocks the task, a needed secret/decision is missing,
  or a verification command fails after one fix attempt: STOP and report
  instead of guessing or widening scope.

A risk-tier example for a rename task, showing the forbidden list doing real work:

## Constraints (hard) — rename `getUser` → `fetchUser`
- Rename only within `src/`; do NOT touch `node_modules`, generated files, or `*.snap`.
- Update all call sites; do NOT change the function body or signature.
- Do NOT add a backward-compat alias unless a call site is outside this repo.

Worked example (filled in)

A vague ask ("add validation to the signup endpoint") becomes a final execution prompt:

# Task: Reject signup requests with an invalid email at api/signup.ts

## Context
- Repo/paths: api/signup.ts, api/__tests__/signup.test.ts
- PM: pnpm   Framework/runtime: Express + TypeScript + Zod (already a dependency)
- Relevant files: api/signup.ts:18 (handler), api/schema.ts (Zod schemas live here)

## Do exactly this
1. In api/schema.ts, add `signupSchema = z.object({ email: z.string().email(), password: z.string().min(8) })`.
2. In api/signup.ts:18, parse the body with `signupSchema.safeParse(req.body)`;
   on failure return 400 with `{ error: 'invalid_input' }` before any DB call.

## Constraints (hard)
- Do NOT: deploy, run migrations, add a new dependency (Zod is already present),
  change the success response shape, or touch any file outside the two listed.
- Preserve: the 201 success response and its body.
- Match existing style (named exports, no default export).

## Verify (must pass before done)
- `pnpm tsc --noEmit`            → 0 errors
- `pnpm test api/signup`         → green (add cases: bad email, short password)
- `pnpm lint api/signup.ts`      → clean

## Acceptance criteria
- [ ] POST with a malformed email returns 400 and never reaches the DB.
- [ ] POST with a valid body still returns 201 with the unchanged shape.

## Report back
- Files changed (path → what), commands run + results, residual risks.

## Stop conditions
- If Zod is NOT already a dependency, or a verify command fails after one fix
  attempt: STOP and report. Do not install packages or widen scope.

Note how the forbidden list ("Zod is already present", "do not touch other files") and the stop condition do the real work: they prevent the agent from npm install-ing a second validation library or "improving" adjacent code.

Common issues & anti-patterns

• Multi-goal prompt. "Fix the bug, add tests, and refactor the module" — three jobs; the agent does each partially. Split into one bounded goal.
• Vague change instructions. "Improve error handling" with no file or behavior named. Specify the file and the exact change.
• Verification by vibes. "Make sure it works" instead of a command with an expected result. Always give the command and the green condition.
• No forbidden list. Omitting "do not add dependencies" invites the agent to npm install its way out of a type error.
• No stop condition. Without one, a blocked agent guesses, widens scope, or fabricates a result.
• Embedded secrets. Pasting a real token into the context block leaks it. Reference the env var name.
• Open-ended scope. "and anything else you notice" guarantees scope creep. Bound the task.
• No output contract. The agent finishes but reports nothing reviewable, so the work cannot be trusted.

Required output

Return the finished execution prompt using the template, ready to paste into a coding agent. It must be self-contained: exact context, exact changes, hard constraints, verification commands, acceptance checklist, output contract, and stop conditions — with zero ambiguity.

Safety

• Bake in the destructive-action prohibitions (no deploys, prod migrations, resets, force pushes, vendor scripts) by default.
• Require reproducible verification, not "looks fine".
• Forbid scope creep and parallel-app/duplicate-repo creation explicitly.
• Reference secrets by env var name only; never embed values.
• Require the agent to stop and report on a blocked or ambiguous state rather than proceed.

Completion criteria

Done means the prompt pins context, lists exact changes plus hard constraints, specifies exact verification and acceptance criteria, defines the report-back contract and stop conditions, and contains no ambiguous instruction.