About Atta
Atta is the set of agents, skills and rules I work with in Claude Code. It started as an open-source framework and grew a layer of my own that keeps several projects in step, remembers what each one learned, and reviews every change before it merges.
atta-devOpen-source framework
- built on, extended: Shared base
Shared baseSkills, core agents, stack detection, templates
- seed and pull (atta-sync): Project
ProjectOne of several, with its own working copy
- push an improvement up: Shared base
- has: Core agents
- generates from its stack: Specialists
- every PR and merge: atta-critic
- session hooks: Publisher on each computer
Core agentsProject owner, architect, code reviewer, librarian
- use: Skills
- read first, file learnings: Project wiki
SpecialistsGenerated team leads and specialists for the project’s stack
- use: Skills
SkillsGrill, review, verify, ship, council, wiki, sync
- hand a job to: Another model
Another modelSecond opinion, council seat
Project wiki
- promote a cross-project lesson: Shared wiki
Shared wikiCross-project lessons, one page per project
- read in the wiki tab: Atta Star control room
atta-criticFresh reviewer, gate on PR open and merge
- pass opens the gate: Pull request on GitHub
Pull request on GitHub
- tickets synced both ways: Atta Star control room
Publisher on each computerTranscripts stay local
- snapshot: cards, wiki pages, working time: Atta Star control room
Atta Star control roomProjects, tickets, wikis, timesheet
Where it comes from
The framework is atta-dev, published on npm and GitHub under the MIT licence by Codelaude. I wrote most of it, with one other contributor. Its idea is that agents guide, review and check the work instead of writing code unasked. The rest of this page, from the shared base to the critic, Atta Star and the timesheet, I built on top of it for my own projects in 2026.
One base, a copy in each project
There is one canonical Atta in a shared folder, and each project keeps a working copy next to its code. A project can fall behind the base but never forks from it. A sync script reports what drifted, pulls the base down, or promotes an improvement from one project up so every project gets it. A check at the start of each session stays silent when a project is in step. I set this up after finding three projects that all claimed version 3.0.0 and matched none of each other.
Agents and skills
Every project gets four core agents: a project owner that routes work, an architect, a code reviewer, and a librarian that keeps rules and knowledge. On the first run, Atta reads the project's stack and generates specialists for it, such as Vue, Fastify or PostgreSQL. About 37 shared skills cover the rest: grilling a plan before code, review, proving a UI change in a real browser, shipping a pull request, and asking another model for a second opinion.
What it remembers
Each project has a wiki: why something is built the way it is, what was tried and dropped, the root cause of a bug. A lesson that holds across projects goes to a shared wiki instead. Sessions read the wiki before searching the code, so each one starts where the last stopped.
The critic
Green tests and a session's own review once missed 17 real defects. Since then every pull request goes past atta-critic: a fresh Claude process with no memory and no idea who wrote the code. It gathers the change itself and treats the PR description as claims to check. A hook refuses to open or merge a pull request until a review of that exact commit has passed. Changes to money or data get two critics, and a few paths, such as deployment and sign-in, only I can merge.
Atta Star and the timesheet
Atta Star is the control room: project cards, a ticket queue across projects, the wikis and the timesheet, behind a single sign-on. Hours come from Claude Code session transcripts by fixed rules: a gap of over an hour is a break, a sign-off ends the day, and minutes are never rounded. The transcripts stay on the computer.
Guard rails
An offline secret scan runs before every commit, with extra rules for Portuguese tax numbers and IBANs. Test data is made up, and a brief for another model is checked for addresses and account numbers before it is sent.