# Implement

> A skill for executing implementation tasks with quality as the top priority. Grounds the request in the codebase, resolves requirements and acceptance criteria, obtains approval for an implementation plan, and applies TDD and Why validation as needed. Trigger on requests like 「これを実装して」「機能を追加して」「バグを修正して」「変更して」「これを作って」 「リファクタリングして」「対応して」 and similar implementation requests. Works in a worktree and asks for user approval via difit before committing. Does not create a PR until the user explicitly requests it. Also runs in a non-interactive autonomous mode when invoked with `autonomous` by an orchestrator (e.g. orchestrate-epic) inside a subagent.

- Skill: `maro114510/implement` (Agent Skill)
- Install (CLI): `npx skillmds@latest add maro114510/implement`
- Raw SKILL.md: https://api.skillmd.com/api/skills/maro114510/implement/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: maro114510 (https://skillmd.com/u/maro114510)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/maro114510/implement

---


# implement

Prioritize implementation quality while auto-adjusting the flow based on task characteristics.
Skip unnecessary steps; go deep only where it matters.

---

## Autonomous Mode

Activated **only when the first token of `$ARGUMENTS` is exactly `autonomous`** — used when an orchestrator dispatches this skill inside a subagent, where no human is reachable.
The word appearing anywhere else (e.g. "implement autonomous reconnection") is task text, not a trigger.
The caller may pass `branch <name>` and `worktree <path>`. Everything not listed below runs as in the normal flow:

- **Before Phase 1**, resolve and validate the worktree. Capture and canonicalize every registered path from
  `git worktree list --porcelain` as `GIT_WORKTREES`. Use the caller-provided path; if none was given, run
  `git wt <branch>` and capture its printed path. After `git wt` succeeds, recapture and canonicalize
  `git worktree list --porcelain` as `GIT_WORKTREES` so the newly created worktree is included. If neither
  value is available, return BLOCKED. If worktree creation fails, return FAILED. Canonicalize the candidate
  with `git -C <path> rev-parse --show-toplevel`,
  then require an exact match in `GIT_WORKTREES`, reject the main checkout (the first `worktree` entry), and
  require a non-detached current branch. When `branch` was supplied, require it to equal
  `git -C <path> branch --show-current`; otherwise set `branch` to that verified value. A missing path, invalid
  repository, unregistered path, main checkout, detached HEAD, or branch mismatch must return FAILED before
  any repository inspection or edit. Never fall back to the current directory or another worktree. When
  blocking or failing before either value exists, emit `UNKNOWN` for its report field. Existing changes in a
  validated worktree are prior work — continue on top of them.
- **Phases 1–2 still run.** Inspect the resolved worktree and treat the Issue body, Epic body, caller prompt, and persisted user answers as the only authoritative product decisions. Do not run an interactive Why check. If a missing Why, requirement, acceptance criterion, constraint, or critical behavior could materially change the implementation, return a BLOCKED report with concrete questions and options. Never guess — the orchestrator relays questions and re-dispatches you with answers.
- **Phase 3**: build the implementation plan, but skip interactive approval only when every material decision is already supported by those authoritative sources. If the plan would introduce an unsupported product or technical decision, return BLOCKED instead.
- **Phase 4**: skip it — the orchestrator already updated the base, and parallel workers would race on the shared checkout. The worktree was resolved before Phase 1.
- **Phase 7**: skip difit — no human to review it. Commit in the worktree with the `commit` skill; never push, never open a PR. Leave the commit there; the orchestrator ships it after human approval.
- **Final output**: exactly this report — it is the return value the caller parses, not a human-facing message:

```
STATUS: DONE | BLOCKED | FAILED
ISSUE: #<number, when the caller supplied one; omit otherwise>
BRANCH: <branch, or UNKNOWN if unavailable>
WORKTREE: <absolute worktree path, or UNKNOWN if unavailable>
HEAD_SHA: <commit SHA of the worker's final commit, or UNKNOWN if unavailable>
CHANGED_FILES: <one path per line; empty if BLOCKED before implementing>
CHECKS: <one per line, `<command> -> exit <code>`>
CRITERIA: <every criterion from Phase 6.9, one per line, `<criterion verbatim from the source> -> <evidence>`; a criterion with no evidence is still listed, with `-> none`>
SKIPPED: <criteria and requirements deferred rather than met, one per line, `<criterion> -> <what was done instead, and why>`; empty if none>
FOUND: <defects found outside scope but not fixed, one per line; empty if none>
SUMMARY: <what was implemented; key decisions and why>
QUESTIONS: <BLOCKED only — numbered, each with concrete answer options>
ERROR: <FAILED only — what failed, what was attempted>
```

---

## Phase 0: Assess the Request

Determine the execution mode and task type from `$ARGUMENTS` and conversation context. Record what the
user has already decided; do not make them repeat it. Defer the TDD decision until after inspecting the
repository and defining acceptance criteria.

In normal mode, do not mutate repository state in Phases 0–3. Autonomous mode has only the worktree-resolution
exception described in Autonomous Mode.

---

## Phase 1: Ground in the Repository

Inspect before asking questions:

1. Read repository instructions and the relevant implementation, tests, types, configuration, schemas,
   documentation, and CI entrypoints.
2. Inspect Git status, the current branch, remotes, and the remote default branch without switching,
   pulling, stashing, or creating a worktree.
3. Summarize the current observable behavior, established conventions, affected interfaces, and constraints.
4. Separate facts discoverable from the repository from genuine product or design decisions. Never ask the
   user for a fact that can be established safely from available sources.

If the repository is unavailable or the relevant source cannot be identified, treat that as a material gap
in Phase 2 rather than inventing an implementation context.

---

## Phase 2: Establish the Implementation Contract

Establish all of the following for every task, using the request, conversation, repository evidence, and
linked specifications:

- **Goal and actor**: who needs what problem solved, and why it matters
- **Current and desired behavior**: the externally observable difference this change must create
- **Requirements**: the behaviors that must be true when the work is complete
- **Specification**: the necessary interfaces, inputs and outputs, state transitions, defaults, precedence,
  and error behavior
- **Scope and non-goals**: what is included and intentionally excluded
- **Acceptance criteria**: concrete, independently verifiable pass/fail outcomes, including relevant failure paths
- **Constraints and compatibility**: supported environments, public interfaces, data or migration obligations,
  performance or operational limits, and prohibited changes
- **Prerequisites and dependencies**: required services, data, permissions, tools, and upstream work
- **Assumptions**: every implementation-relevant belief not guaranteed by a requirement or repository evidence

Run the Why check only when the goal or value is missing and the answer could change whether or what to build.
Ask what problem is being solved, the cost of leaving it unsolved, and whether a smaller change achieves the
same outcome. If the proposed implementation is unnecessary, say so candidly and let the user decide.

### Scrutinize Critical Behavior

For every applicable area, define the required behavior rather than merely noting the risk:

- Authentication, authorization, privacy, secrets, and trust boundaries
- Destructive operations, data integrity, migrations, and backward compatibility
- Transaction boundaries, idempotency, concurrency, ordering, retries, and duplicate delivery
- Validation, partial failure, rollback, cancellation, timeout, and recovery behavior
- Resource limits, performance regressions, observability, rollout, and operational ownership

Surface contradictions between the request and the repository, requirements that cannot all be satisfied,
untestable acceptance criteria, and fatal flaws or logic gaps the user may not have noticed.

### Ask Only Decision-Relevant Questions

Classify each unresolved item:

- Resolve discoverable facts through further inspection.
- State low-risk, reversible defaults as proposed assumptions in the plan.
- Ask about any material decision whose alternatives change user-visible behavior, safety, compatibility,
  scope, architecture, data, or acceptance criteria.

Use `AskUserQuestion` with concrete, mutually exclusive options, a recommended default, and the consequence
of each option. Ask in small groups and continue until no **material** uncertainty remains. Never convert a
security, authorization, data-loss, irreversible, or otherwise potentially fatal gap into an assumption.

In autonomous mode, return a BLOCKED report instead of asking whenever a material decision lacks an
authoritative answer — see Autonomous Mode.

---

## Phase 3: Present and Approve the Plan

Before any repository mutation, present a decision-complete implementation plan containing:

- The goal and observable success criteria
- Relevant current-state evidence
- The implementation approach and affected interfaces or data flow
- Failure, compatibility, migration, and operational behavior where applicable
- A test strategy mapped to the acceptance criteria
- Explicit non-goals, assumptions, risks, and rejected alternatives that materially affect the decision

In normal mode, use `AskUserQuestion` to offer **Approve**, **Adjust**, or **Cancel**. Do not treat answers to
individual clarification questions as approval of the complete plan. If the user adjusts it, update the
implementation contract and present the complete revised plan again. Do not proceed without explicit approval.

In autonomous mode, skip the interactive approval only as described in Autonomous Mode.

If material new information, repository drift, or a scope change invalidates an approved plan, stop, update the
contract and plan, and obtain approval again before continuing.

---

## Phase 4: Set Up the Worktree

In autonomous mode, follow the worktree rules in Autonomous Mode and skip the normal-mode base update below.

For normal mode:

1. Verify the checkout is still clean. Never stash, discard, or overwrite existing changes. If it is dirty,
   stop and ask how the user wants to preserve that work.
2. Resolve the base branch from an explicit approved choice or `refs/remotes/origin/HEAD`. Do not assume `main`.
   If no trustworthy base can be resolved, ask before changing Git state.
3. Switch to the base branch and update it only with a fast-forward pull:
   ```bash
   git switch <base-branch>
   git pull --ff-only origin <base-branch>
   ```
   If either command would overwrite work, diverges, or fails, stop and report the state; do not force it.
4. Choose a GitHub Flow-compliant branch name (e.g., `feat/add-login`, `fix/null-pointer-on-checkout`) and run:
   ```bash
   git wt <branch-name>
   ```

Capture the worktree path printed by `git wt`; its location is configuration-dependent. Run every subsequent
command relative to that path. Because shell state does not persist between tool calls, prefix commands that
need the worktree with `cd <worktree-path> && <command>`.

Recheck the relevant files after setup. If the updated base changed a material premise of the approved plan,
return to Phase 2.

---

## Phase 5: Implement

Choose and record the test approach in the approved plan.

**Proceed with TDD when all conditions are met:**

- The change involves business logic, API handlers, or data transformation.
- A test framework already exists without disproportionate setup cost.
- Inputs and expected outputs can be defined from the acceptance criteria.

**Do not force TDD when any condition applies:**

- The change is limited to UI styling, a migration, configuration, or a one-shot operation.
- No suitable test framework exists and adding one is outside the approved scope.
- Another verification method maps more directly to the acceptance criteria.

### With TDD

Follow t-wada's TDD cycle strictly:

1. **Write the test list** — enumerate all test cases you can think of before writing any code
2. **Red** — write one failing test
3. **Green** — write the minimum code to make it pass
4. **Refactor** — clean up the design while tests stay green
5. Repeat 2–4 until the test list is exhausted

If a test is hard to write, treat it as a signal to revisit the design.

### Without TDD

Implement with the minimum changes. Touch nothing beyond what is required.

---

## Phase 6: Verify CI Locally

Check `.github/workflows/` and `Makefile` / `package.json` for CI configuration, then run the
equivalent checks:
- lint / typecheck
- test suite
- build

Fix any errors before moving on to commit.

---

## Phase 6.5: CodeRabbit Self-Review (conditional)

Select the available CodeRabbit executable and invoke it in the same Bash call, because shell variables do
not persist between tool calls:

```bash
if command -v coderabbit >/dev/null 2>&1; then
  CODERABBIT_BIN=coderabbit
elif command -v cr >/dev/null 2>&1; then
  CODERABBIT_BIN=cr
else
  CODERABBIT_BIN=
fi

if [ -n "$CODERABBIT_BIN" ]; then
  "$CODERABBIT_BIN" review --agent --type uncommitted
fi
```

If `CODERABBIT_BIN` is empty, skip this phase. In normal mode, proceed to Phase 7. In autonomous mode,
skip Phase 7 and emit the required structured report without invoking difit or committing.

1. Run the block above to request a review of the uncommitted diff. If the command errors out (e.g. not
   authenticated, network failure), note this briefly to the user
   and follow the same mode-aware completion path as the missing-tool case: Phase 7 in normal mode,
   or the structured report in autonomous mode.
2. Treat the output as a self-review. Fix only genuine issues, with the minimum change required —
   do not piggyback unrelated cleanup.
   For findings that are false positives or reflect an intentional design decision,
   leave the code as is and note the reason briefly to the user.
3. If any fix was applied, rerun Phase 6 before rerunning the selection-and-review block, so lint, tests,
   and build reflect the fix. Repeat until the review is clean or all remaining findings are judged as not
   requiring action.

---

## Phase 6.9: Reconcile Against the Source

Runs in both modes. Phase 2's criteria live in context, and context drifts toward what you did — so check
against the source again, not against memory.

1. **Re-fetch the source** — Issue body, linked spec, caller's prompt. If it changed since Phase 2, say so first.
2. **Quote each acceptance criterion verbatim.** Paraphrasing is where a criterion gets softened.
3. **Attach evidence**: a command and its exit code, the covering test, a file path, or a quoted user decision.
4. **No evidence means unsatisfied** — a TODO, a new Issue, "out of scope", "future work", a later Epic.
   Deferring a criterion is the user's call; surfacing it is yours.

```
| Criterion (verbatim) | Evidence |
|---|---|
| <the source's own words> | `go test ./...` -> exit 0 |

Deferred, not satisfied: <criterion> -> <what was done instead, and why>
No evidence: <criterion>
```

Lead with those last two lines whenever either has content. In autonomous mode they are what `CRITERIA` and
`SKIPPED` carry.

---

## Phase 7: Request Approval Before Committing

In autonomous mode, skip this phase entirely and emit the structured report — see Autonomous Mode.

Use `difit` to have the user review the diff before committing.
Use `difit` if `command -v difit` succeeds, otherwise use `npx difit`.

```bash
# Review uncommitted changes in the worktree
difit .
```

If review comments come back, address them and run again.
If it exits without comments, treat that as approval to proceed.
Invoke the `commit` skill to compose and make the commit — it decides on its own whether to commit automatically or ask first, independent of the difit review just completed.

**Do not create a PR until the user explicitly says "create a PR."**