// Open source
How I work, in public.
AI coding tools now produce more code in an hour than a person can read in a day. The usual answer is to read every line. That makes the reviewer the bottleneck, and quietly caps how much a team can take on.
Zen Agent Skills is the method I use on client work, released free and open source so you can read it before you hire me. 24 written playbooks that give an AI assistant a consistent way to work. Agree the contract in writing before anything is built. Verify the result with evidence after. Plain Markdown and a few standard-library Python scripts. No platform, no service, no account.
// How it works
Four steps, in order.
-
01
Agree the contract
Before anything is built, the idea becomes a short written specification: what it should do, what it deliberately should not, and how you will know it worked. You approve it. Nothing gets coded against a draft.
In practice: The expensive disagreements happen on one page, while they are still cheap.
-
02
Break it into small jobs
The approved specification is decomposed into atomic tasks, each small enough to finish, check, and undo on its own, each carrying the command that proves it is done.
In practice: Progress arrives in units you can actually review.
-
03
Build in isolation
Independent tasks go to separate agents, each working in its own copy of the project so they cannot trip over each other. Results are reconciled deliberately. Nothing is merged on faith.
In practice: Parallel speed without the parallel mess.
-
04
Verify with evidence
A final independent pass runs the declared checks, audits the result against the specification, and returns pass, fail, or blocked, each with the evidence behind it.
In practice: "It works" becomes a claim someone can check.
It tells you when it cannot be sure.
Most AI workflows report success. Success is the easy thing to report. This one returns one
of three verdicts, and the third is the point. blocked means the check could not
run honestly. Here is a real one from the repository, dated 2026-07-27.
verdict: blocked
blocking_reasons:
- reason: the supplied contract is not approved
detail: docs/spec/house-review.md carries `status: draft`.
A draft spec is one no human has agreed to, so verifying
against it would launder an unapproved contract into evidence.
commands:
- not executed. A blocked run reports no pass or fail for the
work itself, so no verification command was run and none is
reported. That specification has since been approved. The record keeps its original date. A verification that was true last month says nothing about today.
// What's in the box
Seven jobs it does.
- Set up a project and its work tracker so an agent can work in it safely
- Turn a rough idea into a written specification you approve first
- Hold work at the gate until the plan is actually implementable
- Break approved work into small, self-contained, verifiable jobs
- Run several jobs in parallel, isolated, then reconcile them on purpose
- Derive the tests from the approved specification
- Write the docs, the pull request, and the handoff so the next person is not guessing
// Install
git clone https://github.com/hams-ollo/zen-agent-skills.git
cd zen-agent-skills
python scripts/install.py --dry-run The dry run prints exactly what would be installed and where, and changes nothing. Drop the flag when you are ready.
This is the same discipline behind the paid work.
The written specification you approve, the acceptance criteria you can check, the verdict that admits when it cannot be sure. Those are the artifacts, and the repository is where you can read them.