Implementator Lite
A skill pack that makes your coding agent work the same disciplined way on every repo: a plan you approve, tests, a review, a manual test, and a written record of all of it.
Implementator Lite runs in Claude Code and Codex CLI. It is delivered through the Agensi store over MCP, so nothing is installed: your agent reads the skills live, and everything it writes goes into a separate notes repo that belongs to you.
The loop
Every change goes through the same stages, in this order. The agent does the work and stops at the stages marked for you.
- PlanThe agent writes it. You approve it before any code.
- CodeWritten to the repo's own conventions.
- Tests and checksYour test, lint and typecheck commands.
- Self-testThe agent drives the real app or CLI, unhappy paths too.
- ReviewSeverity-ranked. Every finding gets resolved.
- Manual testThe agent writes the steps. You run them.
- CommitOnly after you confirm. You can keep this step for yourself.
- CloseWhen you say so, the work is archived with its record.
Orange marks a stage that waits for you.
Why people use it
- The agent stops guessing your stack. It scans each repo and writes skills for that repo: how to code, test, review and commit there, each rule tied to a file it found.
- Work survives the session. Plans, reviews, test records and a dated journal live in your notes repo, so the next session picks up where this one stopped.
- Many repos, one way of working. Connect as many as you like. Each has its own settings on top of shared defaults.
- Settings change how a stage runs, never whether it runs. Who commits, which browser, which model tier: all yours to set. The plan approval, the review and the manual test cannot be switched off.
What is in the pack
Thirteen skills, a set of rules, and small scripts for the steps that should not depend on an agent's memory. You do not need to memorise any of this: say what you want and the agent picks the skill.
| Skill | What it does |
|---|---|
implementator | Explains the framework and routes you to the right step. |
implementator-setup | Creates the notes repo and asks six questions about how you work. |
implementator-connect-project | Connects a code repo: type, commands, how to test it. |
implementator-generate-skills | Writes that repo's own coding, testing, review and commit skills from what it finds. |
implementator-preflight | Checks a project is ready before work starts. |
implementator-new-initiative | Opens a new feature or ticket in the notes. |
implementator-implement | Runs the loop above. |
implementator-grill | Questions an unclear plan until every decision is settled. |
implementator-diagnose | Works a hard bug: a failing check first, then the cause, then a fix behind a test. |
implementator-mockup | Tries visual options in the running app, shows you screenshots, reverts everything. |
implementator-pr-screenshots | Before and after shots for a pull request. |
implementator-archive | Closes finished work and archives it with its path kept. |
implementator-configure | Shows and changes settings in plain language, and applies presets. |
What stays yours
Everything the skills generate is your own output: the notes repo, plans, journals, test records and the per-repo skills. The pack itself is licensed for live use, so your agent reads it from the store each time and does not save a copy.
What to expect, honestly
- The rules are instructions, not locks. A local install could use git hooks to enforce them. A live pack cannot. One script does refuse to mark work resolved without a plan, a review, a self-test record and a manual run on disk, but an agent can still drift, so read what it reports. If it skips something, here is what to do.
- Tested on a small scale. We ran it from setup to archive in Claude Code and in Codex CLI on Windows, on a small Node command-line project, with the store's delivery emulated. Linux, macOS and other agents have not been tested yet.
- It costs real usage. One full run from setup to a closed feature used about $3.70 on Claude Code. A review on a larger change costs more. The model settings are how you tune this.
- Codex on Windows has one limit you will meet. Its sandbox stops Node from starting child processes, so you run Node test suites yourself. Details.
Next: get started.