Skip to content

wildanfadh/agent-plan-mode-skill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🔒 Agent Plan Mode Skill

Planning-first, edit-later. A reusable skill that brings plan-mode behavior to any AI coding agent or tool — even ones that don't natively support it.

Perfect for vibe coders, solo devs, and teams who want the assistant to inspect, analyze, and propose before touching code, databases, or infrastructure.


License: MIT Status

🎯 What It Does

This skill forces the AI agent into read-only planning mode until you explicitly approve execution. While planning:

✅ Allowed ❌ Forbidden
Read files & directories Edit, create, or delete files
Inspect schemas & configs Run database writes (CREATE, ALTER, etc.)
Run git status, git diff, git log Stage, commit, push, or rebase
Research, search, grep Install/remove/update dependencies
Query DB with SELECT (read-only) Start/stop/mutate services or containers
Dry-run commands Formatters that rewrite files

Once you say "lanjut", "implement", or "approved", the agent switches to full implementation mode following the approved plan.

📦 Quick Install

For pi Coding Agent

# The skill is already in your pi agent directory:
ls ~/.pi/agent/skills/agent-plan-mode/

For Other AI Coding Agents

Copy the SKILL.md file into your agent's skills/prompts directory:

git clone https://github.com/wildanfadh/agent-plan-mode-skill.git
cp agent-plan-mode-skill/SKILL.md ~/your-agent/skills/plan-mode.md

Then add it to your agent's skill library. Most agents auto-discover .md skill files in their skills directory.

Manual Usage

Just paste the content of SKILL.md into your prompt or system instructions at the start of a conversation. The skill's rules will take effect immediately.

🚀 How to Trigger

Start any conversation with one of these triggers:

English Bahasa Indonesia
"plan mode" "mode plan"
"don't implement yet" "jangan edit dulu"
"give me a plan first" "buat rencana dulu"
"read-only analysis" "analisis dulu"
"no changes yet" "read-only dulu"

The agent will:

  1. Inspect the relevant code/files (read-only)
  2. Produce a structured plan with impacted files, steps, risks, and verification
  3. Wait for your explicit approval before making any changes

📋 Example Output

**Tujuan**
Perbaiki bug login: session tidak persisten setelah refresh.

**Konteks Yang Ditemukan**
- Auth middleware di `middleware/auth.ts` tidak menyimpan token ke localStorage.

**Rencana Implementasi**
1. Update `middleware/auth.ts` — simpan token ke `localStorage`.
2. Update `pages/login.tsx` — panggil `auth.persist()` setelah login sukses.
3. Tambah unit test di `tests/auth.test.ts`.

**File/Area Terdampak**
- `middleware/auth.ts`: tambah logika persist
- `pages/login.tsx`: panggil persist
- `tests/auth.test.ts`: test case baru

**Verifikasi**
- `npm test`: semua test pass
- Manual: login → refresh browser → session tetap ada

**Risiko / Pertanyaan**
- Apakah token disimpan plaintext? Pertimbangkan httpOnly cookie.

**Menunggu Persetujuan**
Saya belum akan mengubah file sampai Anda menyetujui rencana ini.

🗂 File Structure

agent-plan-mode-skill/
├── SKILL.md          # The actual skill — copy this into any agent
├── evals/
│   └── evals.json    # Evaluation prompts for testing the skill
├── LICENSE           # MIT License
└── README.md         # This file

🧪 Evaluations

The evals/ directory contains test prompts to validate the skill works correctly across different agents. Use them to ensure the agent respects the planning-only boundary before deploying to production workflows.

🤔 Why This Exists

Many AI coding tools rush into edits. They're great at writing code, but sometimes you want them to think first. This skill creates a safety boundary:

  • Prevents accidental file changes when you just want analysis
  • Improves code quality by forcing the agent to plan before coding
  • Builds trust — you approve the approach before any bytes change
  • Works everywhere — no special agent features needed, just a skill/prompt file

Made for vibe coders who ship fast but want guardrails. 🛡️

📄 License

MIT — see LICENSE for details.

👤 Author

Wildan Fadhgithub.com/wildanfadh


Star this repo if it saves you from premature AI edits!

About

Planning-first, edit-later. A reusable skill that brings plan-mode behavior to any AI coding agent — preventing premature edits until you approve.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors