Implementator

A skill pack for Claude Code and Codex

Your coding agent, with a process.

Implementator is a skill pack for Claude Code and Codex. Every change gets a plan you approve, real tests, independent review and your own manual test. Then a written record, in notes that stay yours.

Get started Watch it work
replay of a real run

The loop

Eight stages. Four of them wait for you.

The same path for every change, on every repo. Pick a stage to see what happens there.

  1. Plan The agent reads the repo and the ticket, then writes a plan. No code until you approve it.
  2. Code Written to the skills generated for this repo, so it follows the conventions already there.
  3. Tests and checks Your own test, lint and typecheck commands, never guessed ones.
  4. Self-test The agent drives the real app or CLI, and forces the failure paths too.
  5. Review Independent subagents, one lens each. Every finding gets resolved, and any fix goes back through the loop.
  6. Manual test The agent writes the steps and the expected results. You run them.
  7. Commit Only after you confirm. Keep this step for yourself if you prefer.
  8. Close When you say so, the work is archived with its plan, reviews and test records.

Proof, not promises

What the loop caught on a demo repo.

Small changes to a small project, and real findings from the reviewers that came after the agent that wrote the code.

4test gaps found in a three-line feature, after its author had called it ready
0pxof layout jump on a slow load, down from 75px, once review reproduced it
8independent reviewers in a full round, one lens each, none seeing the others' findings

Mockups

See it on your real app before anything ships.

The agent tries options on the running page, screenshots each one, and puts the code back exactly as it was. You pick. Your pick goes through the whole loop.

Option A: the existing result line made larger
Option A
Option B: the result shown as two stat tiles
Option B
Option C: the result in a dark band above the text box
Option C, picked
The page after the change, with a dark result band above the text box
The page before the change
BeforeAfter
A merged pull request on GitHub, opened by the agent

Ticket to merge

From Jira to a merged PR, with a paper trail.

  1. Read the ticket. Saved unchanged as the brief, so the plan answers what was asked.
  2. Open a draft PR. Linked to the ticket, in your repo's commit style.
  3. Watch CI. Red goes back through the loop.
  4. You merge. The agent never does. It finds the merge and archives the work.

Fits how you already work.

Skills written for each repoIt scans your code and writes how to build, test, review and commit there. Each rule ties to a file it found. 63 settings, 8 presetsWho commits, which browser, which model tier. Settings change how a stage runs, never whether it runs. Work that survives the sessionPlans, reviews and a dated journal live in a notes repo that belongs to you. Honest about its limitsThe rules are instructions, not locks. Read what it can and cannot promise.

Questions

What people ask first.

What is Implementator?

Implementator is a skill pack for AI coding agents. It makes Claude Code or Codex CLI work through the same eight-stage loop on every repo: plan, code, tests and checks, self-test, review, manual test, commit and close. Four of those stages wait for you.

Which coding agents does it work with?

Claude Code and Codex CLI are the two we have tested, on Windows. Other agents can connect to the store, but the pack has not been tested on them, and Linux and macOS have not been tested yet.

How is it different from a CLAUDE.md or AGENTS.md file?

A rules file tells an agent how your repo works. Implementator writes those rules for each repo from what it finds there, then adds the process around them: a plan you approve, independent review, a manual test, and a record of each change that survives the session.

Do I have to install anything?

No. It is delivered through the Agensi store over MCP, so your agent reads the skills live. You need git and Python 3.8 or newer. gh or glab and a Playwright MCP server are optional.

Can the agent skip the review or the tests?

No setting can switch off the plan approval, the review or the manual test. The rules are still instructions, not locks, so an agent can drift. One script does refuse to mark work resolved without a plan, a review, a self-test record and a manual run on disk.

Does it work with Jira and GitHub?

Yes. It reads a named ticket from Jira or another tracker if your agent has that connector, and it opens pull requests and watches CI on GitHub. It never merges. You do.

Where does my work live?

In a separate git repo of notes that belongs to you: plans, reviews, test records and a dated journal. The next session picks up from there.

How much usage does a run take?

One full run from setup to a closed feature used about $3.70 on Claude Code. Our largest test, from a mockup to a merged pull request with several review rounds, used about $19. The model settings are how you tune this.

Give your agent a process.

Get started Agensi store