> workflow / ai-coding

AI Coding

Implement only approved plan slices inside hooks, stack skills, and architecture boundaries — with tests and docs landing in the same change.

What should I do next?

Complete the steps and evidence below, clear the human gates, then continue to Code Review.

When to use

  • Plan approval HITL has passed for the slice
  • Context package and stack skill are available to the agent
  • pre-code.hook.sh is green for the target slice
  • You are executing, not exploring new architecture

Step-by-step

  1. Load the approved plan slice, ADR IDs, and context package — refuse out-of-plan work.
  2. Select the stack skill (react/node/python/…) and coding.prompt for generation.
  3. Implement the smallest vertical change that satisfies the slice DoD.
  4. Add or update tests with the slice; do not defer "tests later".
  5. Run debug.skill on failures; do not randomly regenerate large diffs.
  6. Update docs/runbooks touched by the change; run post-code hooks.
  7. Open a PR scoped to one slice; link plan item and ADR IDs.

Decisions to make

  • Stop and replan vs push through when reality contradicts the plan
  • Refactor in-slice vs separate debt ticket
  • When to pair-program instead of solo agent execution
  • Whether a spike invalidated assumptions enough to return to human-review

Evidence to collect

  • Feature branch with slice-scoped commits
  • Tests and hook results attached to the PR
  • Plan item ID referenced in PR description
  • No new integrations beyond ADR allowlist

Role-specific next actions

  • PM: Watch for scope creep in PRs; bounce work that is not on the approved plan.
  • Architect: Spot-check boundary violations in early PRs of a new area.
  • Engineer: Drive the agent; own correctness, tests, and hook compliance.
  • QA: Validate test deltas cover the slice acceptance criteria.

Inputs

  • Approved plan slice
  • Context package and stack skill
  • Hook baseline (pre-code / post-code)

Outputs

  • Feature branch / PR
  • Tests and updated docs
  • Hook evidence for the change

Required prompts

  • Coding

    Primary generation prompt constrained by plan and context.

  • Refactoring

    Use only for in-slice refactors with tests — not drive-by rewrites.

Required skills

  • Debug

    Hypothesis-driven failure isolation when generation breaks.

  • Testing

    Keep proof co-located with the change.

  • Documentation

    Update operator-facing docs in the same PR when behavior changes.

Required MCP servers

  • GitHub MCP

    Open/update PRs and read review threads without leaving the harness.

  • Filesystem MCP

    Scoped file ops when the agent runner requires MCP file access.

Tooling & handbook

Recommended agents & patterns

Human approval gates

  • No merge without passing hooks
  • Out-of-plan changes require returning to implementation-planning / human-review

Automation opportunities

  • post-code and pre-commit hooks in local + CI
  • Stack cookbook generate prompts for idiomatic code
  • Budget guards on token spend for long agent sessions

Enterprise best practices

  • Budget-guard.hook.sh for token spend on large slices
  • No secrets in prompts, logs, or agent memory
  • Attribute AI-assisted commits for audit where policy requires

Common mistakes

  • Coding without an approved plan file
  • Disabling hooks to "move faster"
  • Accepting huge unrelated refactors from the agent in a feature PR