Get started

Running in about five minutes.

Pick your agent, copy two commands, then describe the problem in plain language. There is a single runtime and no profile to choose — every host gets the same 26 top-level skills.

Python 3.11+
git
5 supported agents

Six words are enough

You do not need to learn agents, routing or JIT first. These six terms cover everything the rest of this page uses.

Host: The environment that actually provides the model, tools, file system and permissions.
Agent Profile: What one AI worker is responsible for, and which tools it may use.
Professional: Complete judgment for one kind of engineering work — frontend implementation, migration review.
Layer 3: Foundation or Domain capability loaded because this task needs it.
Reference: A file read when a professional question needs deeper judgment.
Evidence: Current source, actual tool output, validation results — and their stated limits.

Step 1

Install it for your agent.

The options below are read from the installer's own tables, so every combination shown is one the CLI accepts. Preview first — step 3 writes nothing — and note that the invocation prefix differs by host.

Which agent do you use?

Where should it be installed?

Available in every repository you open. Start here.

Files land in: ~/.agents/skills
Agent profiles: Native main, analysis, task and review profiles

01Get the repository

git clone https://github.com/machenjie/rd-skills.git
cd rd-skills

02Install the Python package (3.11+)

python3 --version
python3 -m pip install .

03Preview — writes nothing

python3 scripts/quickstart.py --agent codex --scope user --dry-run

04Build, install, run doctor

python3 scripts/quickstart.py --agent codex --scope user

05Re-check any time

python3 installers/doctor.py --agent codex --scope user

06Ask Codex — invocation is host-specific

$engineering-control-plane

Payment callbacks sometimes create the same order twice.
Find the cause and fix it. Add the necessary regression test and verify the change.

Step 2

Describe the problem, not the solution.

Open a small test repository in your agent and say what is wrong. You do not need to investigate first — finding the owner, the tests and the callers is the task agent's job.

A request that is enough on its own

$engineering-control-plane

Payment callbacks sometimes create the same order twice.
Find the cause and fix it. Add the necessary regression test and verify the change.

Add these only if you already know them

A path: If you already know roughly where the problem lives.
Acceptance: How you would recognise that it worked, in one line.
A test command: The command you would run yourself to check it.
A constraint: Anything that must not change — a public API, a schema.

A good request has four parts

None of them needs internal vocabulary. You only need to be asked when a product trade-off, wider scope or extra permission is genuinely your call.

1
Outcome
What to fix, add, or only explain.
2
Constraints
What must not change — a public response, production data.
3
Clues you already have
Paths, errors, a reproducing input, a test command. Clues, not verified facts.
4
Authorisation boundary
Analyse only, or implement and then get an independent security review.

The invocation prefix is host-specific: Codex uses $engineering-control-plane, Claude Code and Copilot CLI use /engineering-control-plane. The wizard above prints the right one for the agent you picked.

Step 3

Know what a good result looks like.

If you get a diff and nothing else, something is wrong. For an implementation request, rd-skills should do all five of these — and an ordinary local change can finish without a separate analysis or review.

it reads the current code before changing it
it finds the owning code and checks nearby consumers
it makes the smallest complete change it can support
it validates after the final edit, not before
it reports changed files, results, limits and any decision still needed from you

A healthy doctor run

Doctor confirms 26 installed skills, the manifest, current source bindings and the agent profile expectation for your host. It proves the files are correct on disk — static checks never prove that a running host loaded them.

If setup fails

Read the conflict before reaching for --force. An old recommended, full or dev manifest means migration: run upgrade rather than uninstalling first.

Troubleshooting and recovery →

Read the report with five questions

Changes: Which files and behaviours changed — and why there?
Validation: Which checks ran after the last relevant edit, and what did they return?
Independent review: When used: which actual change was inspected, and what was found?
Limits: What did not run, was unavailable or flaky, or cannot prove production behaviour?
Open decisions: Is any scope, compatibility, rollout or permission question still yours?

Do not read “tests were written” as “validation is done”. A test existing, a test command running, and a test covering the target behaviour are three different facts.

And what should make you suspicious

Ceremony is the failure mode to watch for. If every small change triggers a planning round and a review, something is routing badly.

Normal behaviour

Behaviour to question

A small change finishes after relevant checks, without another agent reviewing it.
Every change waits for Analysis, a design document, and independent Review.
The Task agent searches for files, tests, and callers itself.
Missing file paths alone cause a separate planning round or a request for you to research the repository.
An unresolved compatibility or authorization decision prompts a focused investigation or question.
Work stops for a vague risk label without a concrete missing fact or affected action.
The original Task repairs an ordinary review defect and reruns targeted validation.
Every repair automatically starts another Review.
An unavailable check is reported with the attempted operation and actual failure.
The result claims success for a check that did not run.

How it works Everyday prompts and troubleshooting Browse the 25 professionals Repository

Reference

Quickstart

The complete document from the repository, including project scope, the OpenAI API zip path and the exact expected outcome.

docs/QUICKSTART.md

This guide starts at a fresh checkout and ends with your first successful rd-skills task. Run commands from the repository root.

1. Install

You need Python 3.11 or newer. Install the repository's declared dependencies:

python3 --version
python3 -m pip install .

Do not copy src/ into an AI coding tool. The setup command builds and installs the supported artifacts for you.

2. Choose your AI coding tool

For a personal Codex installation:

python3 scripts/quickstart.py --agent codex --scope user

Use the matching tool name for another personal installation:

python3 scripts/quickstart.py --agent claude --scope user
python3 scripts/quickstart.py --agent copilot --scope user
python3 scripts/quickstart.py --agent cline --scope user

To keep rd-skills inside one project, provide that project root:

python3 scripts/quickstart.py --agent codex --scope project --target /absolute/path/to/project

Preview the exact plan without building or writing files:

python3 scripts/quickstart.py --agent codex --scope user --dry-run

Add --verbose when you need the command plan and detailed command output.

Supported hosts are codex, claude, copilot, cline, and openai-api. Codex supports project, user, and admin; Claude, Copilot, and Cline support project and user. OpenAI API produces zip files and has no installation scope.

For OpenAI API packages, run:

python3 scripts/quickstart.py --agent openai-api

For Codex admin scope, non-default targets, exact installed paths, and permission boundaries, use Advanced Installation & Recovery.

3. Verify

Quickstart checks the installation automatically for Codex, Claude, Copilot, and Cline. You can run the check again:

python3 installers/doctor.py --agent codex --scope user

A healthy result is intentionally short:

✓ rd-skills installed
✓ expected configuration found
✓ installation healthy

Next:
Open or restart Codex.
Start with $engineering-control-plane and describe the task in natural language.
The full rd-skills workflow is available on Codex.

Use --verbose to see inventory, source binding, host configuration, and digest details:

python3 installers/doctor.py --agent codex --scope user --verbose

Doctor verifies installation artifacts. It does not prove your AI coding tool loaded rd-skills. Open or restart the tool before the first task.

4. Run your first task

Open a small repository in your AI coding tool. If the table below gives a verified invocation for that host or surface, start the request with it.

Host invocation

Host or surface	Artifact delivery	Live Skill invocation	Full rd-skills workflow	Limit
Codex	Skills + Agent Profiles	`$engineering-control-plane`	Available	Artifacts checked; live loading not proved
Claude Code	Skills + Agent Profiles	`/engineering-control-plane`	Available	Artifacts checked; live loading not proved
Copilot CLI	Skills + Agent Profiles	`/engineering-control-plane`	Available	Copilot CLI only
Cline	Skills only	Not established	Not established	Artifact delivery only
OpenAI API	Zip packages	Not applicable	API integration owns orchestration	API integration only

Other Copilot surfaces are separate surfaces. Do not assume the Copilot CLI invocation or workflow availability applies to them.

For Codex, enter this normal engineering request:

$engineering-control-plane

Payment callbacks sometimes create the same order twice.
Find the cause and fix it. Add the necessary regression test and verify the change.

Do not type /engineering-control-plane at the start of a Codex request. Codex reserves leading Slash input for its own command menu and uses the $skill-name form for Skills.

You do not need to know the responsible file, the exact test, or rd-skills internals. If you already know useful facts, add them below the request:

The callback handler is probably under payments/.
Keep the public response unchanged.
The repository test command is pytest -q.

rd-skills should inspect the current code, identify the cause and affected boundary, make a bounded change, run relevant checks after the final edit, and report what changed and what remains unverified. Deeper Analysis is used for unresolved decisions; independent Review is used when requested or justified by a concrete risk.

Continue with Usage for everyday task examples and result interpretation.

5. If it did not work

If quickstart failed, rerun the same command with --verbose and keep the first specific error.
If doctor reports a file, configuration, permission, ownership, or migration problem, follow Troubleshooting and Recovery. Do not add --force until you have inspected the exact conflict.
If doctor is healthy but the tool does not respond to the invocation in the table, restart the tool and confirm you installed for the same tool and scope. Doctor does not prove that a live host loaded the files.
In Codex, type $engineering-control-plane, not /engineering-control-plane.
For other Copilot surfaces, confirm that surface's current Skill invocation support instead of assuming the Copilot CLI syntax applies.
For Cline, doctor can verify installed files, but the current host contract does not establish live Skill loading, an explicit invocation, validation, or independent review.
If the task stops for a decision, answer the concrete scope, safety, compatibility, or production question instead of broadening the request implicitly.
For a reproducible problem, use Support and remove secrets from the output you share.