plan-implement-review

name

plan-implement-review

description

End-to-end feature workflow from user instructions to shipped code. Plans with an Opus subagent in Plan mode (read-only exploration, no implementation). If the plan has open questions, obtains user answers/approval; otherwise proceeds immediately. Then runs the implement-review skill (dev subagent implements, Opus reviews, dev fixes, orchestrator summarizes). Use when the user gives a feature request or task and wants planning plus implementation with review. Trigger on "plan, implement and review", full feature delivery from instructions, or when no approved plan exists yet.

Plan, Implement, Review

Overview

This skill extends implement-review with an upfront planning phase:

1. Plan — Opus subagent explores the codebase in Plan mode and produces a structured plan.
2. Gate — If the plan has open questions for the user, resolve them and get approval. If it has none, skip validation and proceed.
3. Implement → Review → Fix → Summarize — Follow the implement-review skill with the (auto-)approved plan.

The orchestrator coordinates subagents and writes the final summary. It does NOT plan, implement, or review code itself.

Preconditions (verify before any work)

Stop and ask the user if either is missing — do not start.

1. User instructions exist. The user has stated what they want built or changed. If vague, ask clarifying questions before launching the planning subagent.
2. The orchestrator is running the auto model in Agent mode. State which model and mode you are on. If not auto in Agent mode, stop and ask the user to switch before continuing.

Workflow checklist

- [ ] Step 0: Preconditions verified (user instructions + orchestrator on `auto` in Agent mode)
- [ ] Step 1: Plan in an Opus subagent (Plan mode — read-only)
- [ ] Step 2: Present plan; ask for approval only if there are open questions
- [ ] Step 3–6: Follow implement-review skill (implement → review → fix → summarize)

Step 1: Plan in an Opus subagent (Plan mode)

Launch a planning subagent that behaves like Cursor Plan mode: explore and design only, no code changes.

• Tool: Task with subagent_type: "explore" (preferred for codebase exploration) or subagent_type: "generalPurpose", readonly: true.
• Model: claude-opus-4-8-thinking-high (Opus). If that slug is unavailable, tell the user Opus is unavailable rather than silently substituting another model.
• Prompt must include:
  • The full user instructions (verbatim or faithfully summarized).
  • Instruction to act in Plan mode: read files, search the codebase, understand architecture (AGENTS.md, .cursor/ARCHITECTURE.md), and produce a plan — do not write or modify any files.
  • Required plan output format:

# Plan: [short title]

## Goal
[One paragraph]

## Deliverables
1. [Concrete, checkable deliverable]
2. ...

## Approach
[Key technical decisions, files to touch, patterns to follow]

## Risks / open questions
- [Anything needing user input, or "None"]

• Ask the subagent to return the complete plan and any open questions. Instruct it to put "None" under Risks / open questions when nothing needs user input (non-blocking risks alone do not count as open questions).

Record the planning subagent's agent ID if the user requests plan revisions — resume it with feedback instead of starting over.

Step 2: Present plan; approval only when there are open questions

1. Present the plan to the user clearly (the numbered Deliverables list is what implement-review will track).
2. Decide whether the plan needs user validation:
   • Open questions present (anything under Risks / open questions that needs a user decision or input — not "None"): resolve those questions with the user, then wait for explicit approval ("go ahead", "approved", or equivalent) before continuing. If they want plan changes, resume the planning subagent with their feedback and repeat this step.
   • No open questions (Risks / open questions is "None", empty, or only non-blocking notes that do not need a user answer): do not ask for approval. Treat the plan as approved immediately, briefly note that you are proceeding because there were no open questions, and continue to Steps 3–6.

Once the plan is approved (explicitly or by auto-proceed), restate the deliverables as the approved plan for the next phase.

Steps 3–6: Implement, review, fix, summarize

Read and follow implement-review/SKILL.md using the approved plan from Step 2.

• The implement-review preconditions are satisfied once the plan is approved (including auto-proceed with no open questions) and you remain on auto in Agent mode.
• Start at implement-review Step 1 (dev subagent implementation). Do not re-run planning.
• The final summary (implement-review Step 4) is the deliverable of this skill.

Notes

• Planning subagent: read-only, Opus, no implementation.
• Implementation and fix subagent: auto model (inherit from orchestrator), same subagent reused for fixes.
• Review subagent: read-only, Opus, angular-code-review skill.
• If the user already has an approved plan and only wants implementation, use implement-review directly instead of this skill.