rocket-stack · open source · MIT

Your AI coding agent says “done”. Make it prove it.

rocket-stack is an open-source skill pack for AI coding agents. Before your agent calls a task done, it writes a short test plan, reviews its own diff, runs the feature for real (UI, API or CLI) and saves the evidence in your repo.

Tested with Cursor · Claude Code Should work with any agent that supports SKILL.md skills.

› /rocket-mode

› add login rate limiting

Playbook: Ship → implement, review, prove

# Mission Log

Playbook: Ship

Next: /pre-flight-check login → /mission-control login

Result: Pass

Proof: mission-map/login/proof/pre-flight.md

mission-map/login/proof/dashboard.png

Note: Rate limit added; Pre-Flight Check Go; login Test Flight passed.

The problem

Agents often mark work as done without showing that it works.

The fix

rocket-stack makes proof part of “done”: no Test Flight, no proof, not done.

How it works

One mission. Five stages.
Proof at the end.

Turn on Rocket Mode and describe the task like you would to a teammate. It picks the right playbook and routes every change through the same five stages.

  1. 01 · Map

    Mission Map

    A short test plan per feature: how to reach it, how to control it, and what proof looks like. Kept in sync with the code.

    /create-mission-map
  2. 02 · Build

    Ship the change

    You type the task in plain language. Rocket Mode picks the playbook and hands the work to the right skills.

    /rocket-mode
  3. 03 · Review

    Pre-Flight Check

    The agent reviews its own diff like a strict peer. Six checks, every finding with file:line. Go or No-go.

    /pre-flight-check
  4. 04 · Fly

    Test Flight

    The agent runs one real user action on the real surface (a UI, an API or a CLI) to prove it works.

    /mission-control
  5. 05 · Prove

    Mission Log

    Proof saved in your repo: the action, the result and the side effect. Pass or Fail, with the path.

    mission-map/<name>/proof/
Words you’ll see
Rocket Mode
The entry point. Describe a task; it picks the playbook and the skills.
Mission
One feature or user flow, like login or signup.
Mission Map
A short test plan for one mission: how to reach it, control it and prove it.
Pre-Flight Check
The agent’s strict review of its own diff. Every finding cites file:line.
Go / No-go
The review verdict. Any open finding is a No-go and the agent stops.
Test Flight
One real user action on the real surface: UI, API or CLI.
Mission Log
The saved proof: action, result and side effect, with Pass or Fail.
Ground Control
A repo-wide quality scan you can run anytime, outside the five stages. Its findings list is called Space Debris.
The skill pack

Six skills. One standard of done.

Start with Rocket Mode and let it route, or call any skill directly. Each one writes only to its own folder and reports in a fixed format.

Start here

Rocket Mode

/rocket-mode

Turn it on and describe the task in plain language. It picks a playbook (Map first, Prove, Drift, Ship, Fix fail or Ground control), routes to the other skills and insists on proof before “done”.

Writes & reports
Writes
Nothing itself. It routes.
Reports
Playbook · Next · Result · Proof · Note

Mission Control

/mission-control <name>

Proves one mission: updates its Mission Map if needed, runs a Test Flight and leaves a Mission Log. On a fail it still captures proof when possible.

Writes & reports
Writes
mission-map/<name>/
Reports
Result (Pass / Fail) · Proof · Note · Updated

Create Mission Map

/create-mission-map

Interviews the repo, writes the top 3–5 missions (happy paths first, then auth, then fragile flows) and proves one with Mission Control.

Writes & reports
Writes
mission-map/
Reports
Created · Proven · Proof · Note

Maintain Mission Map

/maintain-mission-map

Fixes drift when routes, commands, selectors or flows change. Keeps every existing proof file and lists the missions that need a new proof.

Writes & reports
Writes
mission-map/
Reports
Changed · Reason · Re-prove

Pre-Flight Check

/pre-flight-check <name>

Reviews its own diff like a strict peer, before the Test Flight. Brief, Tests, Errors, Security, Readability, Debris: every finding with file:line. Any Open means No-go.

Writes & reports
Writes
mission-map/<name>/proof/pre-flight.md
Reports
Result (Go / No-go) · Fixed · Accepted · Open · Proof · Note
Anytime

Ground Control

/ground-control

Repo-wide, outside the five-stage flow: run it anytime. Scans for duplicates, dead code, missing tests, oversized files, fragile code and debris, keeps a Space Debris list with a trend and picks the top 1–3 to fix next. It never changes app code.

Writes & reports
Writes
ground-control/debris.md
Reports
Open · Trend · Next · Proof
Proof, not promises

Every claim leaves a trail in your repo.

Maps and proof live in the target repo, next to the code, not in the skill pack. Anyone on the team can open them and check.

Plain files, reviewed like code

Markdown, screenshots and command output under mission-map/. Commit them and they show up in code review like any other change.

Evidence or it isn’t a finding

Review findings and debris items cite file:line. No evidence, no item.

History is the trend

Proof is never deleted, and Space Debris items are never removed, only marked Fixed or Accepted.

your-app / target repoexample files
mission-map/
login/
proof/
ground-control/
Example run · your-app /rocket-mode
  1. › /rocket-mode
  2. › add login rate limiting
  3. Playbook: Ship → implement, review, prove
  4. ✓ Implemented in src/auth/login.ts
  5. /pre-flight-check login
  6. ✓ Pre-Flight Check: Go · 2 fixed · 0 open
  7. wrote mission-map/login/proof/pre-flight.md
  8. /mission-control login
  9. Test Flight · UI
  10. Enter email and password, click Login
  11. ✓ Result: Pass redirected to the dashboard
  12. saved mission-map/login/proof/dashboard.png
dashboard.pngExample screenshot, developed after the Test FlightResult: Pass
Quick start

Ready for launch in two steps.

No build step and no config. Each skill is a plain folder with a SKILL.md your agent reads.

  1. Install the skills

    Cursor

    Clone the repo and copy the six skill folders into your personal Cursor skills folder, so they work in every project:

    git clone https://github.com/rocketbase-space/rocket-stack.git
    cd rocket-stack
    mkdir -p ~/.cursor/skills
    cp -R rocket-mode mission-control create-mission-map \
          maintain-mission-map pre-flight-check ground-control \
          ~/.cursor/skills/

    One project only? Copy them to .cursor/skills/ in your repo root instead and commit them, so your team gets them too.

    Claude Code

    Clone the repo and copy the six skill folders into your personal Claude Code skills folder, so they work in every project:

    git clone https://github.com/rocketbase-space/rocket-stack.git
    cd rocket-stack
    mkdir -p ~/.claude/skills
    cp -R rocket-mode mission-control create-mission-map \
          maintain-mission-map pre-flight-check ground-control \
          ~/.claude/skills/

    One project only? Copy them to .claude/skills/ in your repo root instead and commit them, so your team gets them too.

    Other agents

    Copy the six skill folders into the directory your agent loads skills from. Each folder holds one SKILL.md.

    git clone https://github.com/rocketbase-space/rocket-stack.git
    cd rocket-stack
    cp -R rocket-mode mission-control create-mission-map \
          maintain-mission-map pre-flight-check ground-control \
          <your-agent-skills-dir>/

    Not tested yet, but it should work with any agent that supports SKILL.md skills. Check your agent’s docs for its skills folder.

  2. Launch

    Enable Rocket Mode, then type the task as a normal message. No flags, no forms.

    › /rocket-mode
    › add login rate limiting

    Or call a skill directly:

    /mission-control login /create-mission-map /maintain-mission-map /pre-flight-check login /ground-control

What you need

  • An agent that loads SKILL.md skills. Tested with Cursor and Claude Code.
  • A git repo. Pre-Flight Check reviews your diff, untracked files included.
  • An app it can run. It reuses a running app or finds the start command (package.json scripts, Makefile, README or Dockerfile), asks once if none is obvious, and stops what it started.
  • For UI proof, a way to take a screenshot, such as a browser tool for your agent. API proof is the response; CLI proof is the output.

What changes in your repo

  • mission-map/<name>/README.mdOne Mission Map per mission, in Markdown.
  • mission-map/<name>/proof/Screenshots, responses or output, plus pre-flight.md (replaced on each run). Proof is never deleted.
  • ground-control/debris.mdOnly when you run /ground-control.
  • Your app codeOnly for a task you asked for, like building or fixing something. The other skills write only to their own folders, and never to the skill files.

It asks at most three questions per run; otherwise it notes its assumptions and carries on.

rocket-stack · MIT

The mission succeeds
only when proven.

Give your agent the same standard you would give a teammate: show it works, then call it done.