/feature

name: feature
description: >
  Runs the full AI-assisted workflow for a non-trivial feature or change:
  plan, plan critique, user approval, implementation with tagged tests, code
  review, exploratory QA, and PR. Use when the user asks to build or start a
  feature, or to implement a change that needs a plan, review, and PR. Every
  planning run saves the approved plan to a file; with `--plan-only` it stops
  there, and `--from-plan <slug>` runs such a saved plan without re-planning.
license: MIT
metadata:
  author: André Cunha

/feature

Pipeline: plan → critique → implement → test → code-review → QA → PR. Run steps in order, skipping only where a step or the mode says so. `AGENTS.md` is the source for commands, default branch, sensitive areas, and task tracking.

`planner`, `plan-critic`, `code-critic` and `adversarial-qa` below are sub-agents: "call" means spawn the sub-agent of that name, never load the same-named skill into your own context (you would be reviewing your own work). The same holds for `build.md` step 4's second reviewer, a `general-purpose` sub-agent that runs the bundled `/code-review` skill.

Input

• User prompt (verbatim, including any doc/link references)
• Session branch: a feature branch, not the default branch (any branch with

`--plan-only`)

Phases

The steps live in files in `${CLAUDE_SKILL_DIR}/`, the directory of this skill. Read each with `Read` when its phase starts:

• `plan.md`: step 1 (1a–1d), from prompt to saved plan
• `build.md`: steps 2–8, from implementation to a reviewed, QA'd HEAD. Its

input is the approved plan text from step 1, or a saved plan it loads itself (`--from-plan`)

• `pr.md`: step 9, opening the PR and keeping it up to date once open

The mode is set by the first token of the arguments; a flag anywhere else is prompt content, passed on verbatim:

| First token | Runs | |---|---| | `--plan-only` | `plan.md`, then STOP | | `--from-plan` | `build.md` (loading the saved plan) → `pr.md` | | anything else | `plan.md` → `build.md` → `pr.md` |

Only this table decides what runs next: each phase file ends at its last step, then continue with the next file in the table. For every phase file:

• An unexpanded `${CLAUDE_SKILL_DIR}`, or a phase file you can't read → STOP

per `AGENTS.md` Rule 2; never work from memory of it

• Re-read the current phase file whenever it is no longer fully in your

context (e.g. after compaction), and read a file before following a pointer into it

• Step numbers are the same across files: steps 1a–1d are in `plan.md`,

steps 2–8 in `build.md`, step 9 in `pr.md`

`--plan-only` STOP: skip steps 2–9, report the plan file's absolute path and that `/feature --from-plan <slug>` on a feature branch runs it. No branch, commit, push, or PR.

Shared definitions

Used by more than one phase file.

**Plan file location**: `.workflow-plans/<slug>.md` under the **main worktree**: the parent of the `.git` dir printed by `git rev-parse --path-format=absolute --git-common-dir`. That parent still contains a `.git` segment (e.g. a submodule), or the dir's name is not `.git` (a bare repo) → ask the user where to write (for `--from-plan`, where to read).

**Ownership check**, before this run rewrites, removes or marks a plan file: the file still has the `slug`, `base-commit` and exact content this run last wrote to it, or read from it for `--from-plan` (when marking consumed, a `consumed:` line is not part of the comparison). Content no longer available, e.g. after context compaction → can't verify → ask.

Output

• Default and `--from-plan`: `pr.md`'s *Output*
• `--plan-only`: `plan.md`'s *Output*, reported as in the `--plan-only` STOP
• Every mode: a summary of skipped steps, plan deviations, `context7` lookup

failures, and why

Rules

• Never skip `planner` for a new feature (`--from-plan` runs a plan that

already went through it)

• Never skip `plan-critic` unless all step 1b criteria hold AND the user opted

out (`--from-plan` re-runs it only when the base moved)

• Never leave a source the prompt references unread: read it yourself and

pass it to the `planner` verbatim; the planner never fetches

• Never enter plan mode from a sub-agent
• Never present work to the user before `code-critic` has reviewed it
• Relay ALL `NEEDS_DECISION` items and Open Questions to the user; never

decide them

• Sub-agents never modify the project's code; only the main agent implements

changes (`adversarial-qa` also writes `.qa-evidence/`)

<!-- SKILL NAMING NOTE (Claude Code): the skill is named `code-critic` so it does NOT shadow Claude Code's bundled `code-review` skill (project skills shadow bundled ones). `/code-review ultra` stays reachable as an optional, user-launched deep pass. Do not rename back. -->