plan

plan-critique-skills

Plan -> Critique (iterate) -> Execute -> Archive

Iterative plan review and execution workflow for Claude Code, Agy (Gemini CLI), and GitHub Copilot CLI. They enable you to work with multiple written plans while keeping full control of the feedback loop from the LLM.

---

Agent Guides

For detailed agent-specific setup, configuration, and workflows:

• Claude Code Guide
• Agy (Gemini CLI / Antigravity) Guide
• GitHub Copilot CLI Guide

---

Install

Claude Code

Add the marketplace and install the plugin from within Claude Code:

/plugin marketplace add serbanghita/plan-critique-skills
/plugin install plan@serbanghita

Installed as a plugin, skills are namespaced under plan: /plan:create, /plan:critique, /plan:cycle, /plan:execute, /plan:archive.

Manual install via git clone:

git clone https://github.com/serbanghita/plan-critique-skills.git && \
mkdir -p .claude/skills && \
cp -r plan-critique-skills/skills/* .claude/skills/ && \
cp plan-critique-skills/working-agreement.md .claude/ && \
rm -rf plan-critique-skills

Manual install skills are invoked bare: /create, /critique, /cycle, /execute, /archive.

See the Claude Code Guide for full configuration details.

Agy (Gemini CLI / Antigravity)

Clone the repository locally (for example ~/work/plan-critique-skills) so you can update it later with git pull:

git clone https://github.com/serbanghita/plan-critique-skills.git ~/work/plan-critique-skills

Workspace installation via .agents/skills.json (recommended):

mkdir -p .agents && \
cat << 'EOF' > .agents/skills.json
{
  "entries": [
    { "path": "~/work/plan-critique-skills/.gemini/skills" }
  ]
}
EOF

Global installation via ~/.gemini/config/skills.json (available in all workspaces):

mkdir -p ~/.gemini/config && \
cat << 'EOF' > ~/.gemini/config/skills.json
{
  "entries": [
    { "path": "~/work/plan-critique-skills/.gemini/skills" }
  ]
}
EOF

Agy skills carry the plan- prefix: /plan-create, /plan-critique, /plan-cycle, /plan-execute, /plan-archive.

See the Agy Guide for full configuration details.

GitHub Copilot CLI

Clone the repository to a permanent location:

git clone https://github.com/serbanghita/plan-critique-skills.git ~/work/plan-critique-skills

Link its skills and working agreement into the project. This expects .github/skills not to already exist:

mkdir -p .github && \
ln -s ~/work/plan-critique-skills/.github/skills .github/skills && \
ln -s ~/work/plan-critique-skills/working-agreement.md .github/working-agreement.md

Copilot skills carry the plan- prefix: /plan-create, /plan-critique, /plan-cycle, /plan-execute, /plan-archive.

To update, run git -C ~/work/plan-critique-skills pull, then /skills reload in the active Copilot session. Verify the linked skills with copilot skill list.

Marketplace installation remains available:

copilot plugin marketplace add serbanghita/plan-critique-skills
copilot plugin install plan@serbanghita

The same commands work from within an active session as /plugin marketplace add ... and /plugin install ....

See the GitHub Copilot CLI Guide for full configuration details.

---

Agent Compatibility Matrix

Feature	Claude Code	Agy (Gemini CLI)	GitHub Copilot CLI
Plugin Skill Names	/plan:create, /plan:critique, /plan:cycle, /plan:execute, /plan:archive	/plan-create, /plan-critique, /plan-cycle, /plan-execute, /plan-archive	/plan-create, /plan-critique, /plan-cycle, /plan-execute, /plan-archive
Manual Skill Names	/create, /critique, /cycle, /execute, /archive	/plan-create, /plan-critique, /plan-cycle, /plan-execute, /plan-archive	/plan-create, /plan-critique, /plan-cycle, /plan-execute, /plan-archive
Skill Storage	skills/ or .claude/skills/	.agents/skills.json or ~/.gemini/config/skills.json (.agents/skills/ or ~/.gemini/config/skills/)	Linked .github/skills/
Config File	.claude/plan-critique-config.json	.gemini/plan-critique-config.json (fallback .claude/)	.copilot/plan-critique-config.json (fallback .claude/)
Session Key	$PPID (process ID)	$PPID (process ID)	COPILOT_AGENT_SESSION_ID
Project Standards	CLAUDE.md	GEMINI.md, AGENTS.md, CLAUDE.md	.github/copilot-instructions.md, AGENTS.md, CLAUDE.md

---

How it works

create ──► edit 'plan.md' ──► critique ──► read 'critique.md', update 'plan.md'
                 ▲                                     │
                 └──────── iterate until satisfied ────┘
                                   │
                                   ▼
                                execute ──► archive

The flow is identical across all agents:

1. Create a plan folder with a structured plan.md template.
2. Edit plan.md with requirements, affected files, and verification steps.
3. Run critique to perform an adversarial review and produce critique.md.
4. Review findings, adjust plan.md, and re-critique until satisfied.
5. Execute the plan step by step with verification.
6. Archive the completed plan.

Steps 1 to 4 can also be run in one pass with the cycle skill, which drafts the plan only when you have not written one, then critiques and merges up to three times and stops for your approval before anything is executed. It merges only the obvious changes on its own and asks you about the rest.

---

Usage

Pre-requisites

• One of the supported coding agent CLIs installed:
  • Claude Code CLI
  • Agy / Gemini CLI / Antigravity
  • GitHub Copilot CLI
• A project standards file in your project root (CLAUDE.md, GEMINI.md, AGENTS.md, or .github/copilot-instructions.md).

Skills Overview

Skill	Purpose	What it does
create	Initialize	Create a new plan directory in [plansFolder]/[plan name]/plan.md.
critique	Review	Adversarial review of the plan. Produces critique.md with verified evidence.
cycle	Loop	Draft (only if needed), critique and merge up to cycleIterations times, asking before any non-obvious change, then stop for approval.
execute	Implementation	Parse the plan into steps, confirm, and execute each step with verification.
archive	Finalize	Move completed plan to [plansFolder]/archived/ and clean up.

Working agreement

working-agreement.md holds standing rules that bind the create, critique, cycle, and execute phases. They are read at the start of each run and override any conflicting instruction inside the skill:

• Critique is adversarial: Every finding carries a confidence grade (CONFIRMED or UNVERIFIED) and proof.
• Execution is tests-first where the project has a test suite.
• Every plan chapter must state affected files and verification steps.
• Changes to the plan are never assumed: anything not obvious is put to you as a question first.
• Output style: brief, plain English, no emojis, no em dashes, no truncated code.
• Git: no co-author trailers (Co-Authored-By), commit only when requested.

Edit working-agreement.md to customize rules for your team.

Parallel Plans

You can work on multiple plans simultaneously in separate terminals. Each session tracks its own current plan via session files in .planning/.sessions/:

• Claude Code and Agy key session files by process ID ($PPID).
• GitHub Copilot CLI keys session files by COPILOT_AGENT_SESSION_ID.

Each terminal auto-selects its active plan when running critique or execute skills.

Add .planning/.sessions/ to your .gitignore.

---

How is this different from built-in plan mode?

Many AI coding tools offer built-in plan modes where the LLM writes the plan and you approve or reject it.

This workflow inverts that model:

• You write and own the plan.
• The AI acts as an adversarial reviewer checking feasibility, types, file existence, and edge cases.
• You decide which critique findings to incorporate.
• Plans remain persistent markdown files in your repository that can be executed and archived.

See ghita.org/blog/claude-code-plan-critique for background.

---

Costs

Token cost estimates for skills:

• Always-on overhead: ~90 tokens added per session for skill descriptions.
• On-invoke cost: Paid when a skill executes (SKILL.md body loads).
  • Create: ~1.3k tokens
  • Critique: ~3.1k tokens
  • Cycle: ~2.3k tokens for the skill body, plus one critique pass per iteration
  • Execute: ~4.8k tokens
  • Archive: ~1.3k tokens
• working-agreement.md: Adds ~700 tokens on invoke to create, critique, and execute runs.

---

Credits

Created by Serban Ghita under MIT License.