Your AI agent says “done”. Is it?
A working discipline for people who run projects that AI agents build: how to know the work is finished when you can't read the code.
Three times in one project, a tool reported success over work it had not done
- A renderer printed a clean log, with timings for every frame and the name of the output folder. The folder existed. It was empty.
- An analytics script recorded 0 views for a video, next to an average watch time of 1 minute 20 seconds. It had looked for a column that wasn't there, and quietly wrote a zero.
- The same script left one of our posts out of every report for four days. No error. It finished cleanly.
None of them crashed. Every one said “done”.
In vibration analysis, a machine that runs is not a machine that's healthy. In administration, a delivery is complete when there is a signed delivery note that someone who wasn't there can check. Running a project with an AI agent needs both habits. The agent writes the code; you make the decisions and ask for the delivery note.
The Agent Discipline Kit
A handful of plain-text files your agent reads at the start of every session and follows while it works. No programming and no terminal needed.
- A project template that interviews you in the first session and fills itself in.
- One log as the single source of truth, so the next session starts from paper, not from memory.
- A handover at every break: what was done, what wasn't, who does what next.
- Ask, don't guess: the agent asks for missing input and says out loud when it is assuming.
- A five-minute checklist for the person in charge.
Install it by pasting one message to your agent:
Install the kit from https://github.com/kastalis/agent-discipline-kit into this folder, then follow AGENTS.md. See the kit on GitHubOwner's Edition
The free kit says what to do. The Owner's Edition is how we actually did it, with the tools and records from a real project.
- The check that found the faults: a small tool that compares what your documents claim with what is actually on disk, with a guide to adapting it to your project.
- Templates: a decision register, a cost log, and a briefing for an expensive review session.
- A case study: one month of a real project, cleaned up, decision by decision.
- The full story: which failure produced which rule.
Planned price: $29, one payment, for use in all your own projects and work.
It isn't on sale yet. Leave your email and you'll get one message when it opens.
About
I'm a mechanical engineer and a vibration analyst, and I have also spent a large part of my career in administration: handovers of duty, a quality system, contracts and tenders, committee work, investments. Now I run projects that AI agents build. I don't write the code; I make the decisions and check the delivery notes.
Kastalis is the studio those projects are published under.
This is not a way to make money with AI, and nothing here promises income. It is a record of what one project needed in order to trust its own reports.