Skip to content
stiglo
NO

README · HOW I WORK

How I work

AI agents make me faster, but they do not get the last word. I work in a fixed flow where every step has an owner and a clear boundary.

goals: defined and written down before every session
guardrails: what the agent must not do, explicit and enforceable
human_review: decisions, direction and spot checks stay with a human

The flow

Work moves forward when the criteria are green, and back when they are not. The flow itself also changes along the way, as I learn what works.

  1. Context and requirements

    Goals, boundaries and success criteria are written down before anything is built. Requirements can change along the way, but as a documented decision.

  2. Tools

    I build my own skills for Claude, for the job and the domain the product lives in, plus persistent rules the agents must follow, for documentation and design principles among other things. What gets checked and enforced depends on the context.

  3. The work loop

    The agent builds in small, bounded deliveries: attempt, validate against the criteria, correct.

    all green?
    no correct and rerun ↺ 3

  4. Checkpoint

    This is where the human sits: direction, trade-offs and spot checks. New insight sends the flow back, not forward.

    no tools improved → 2
    approved?
    no requirements adjusted → 1

  5. Handover

    The state is written down: what is done, what remains, and the next step. The next session, or a completely new agent, can take over without any verbal context.

    ⟲ next session, with full history → 1

Guardrails in practice

This site is itself built this way. The requirements live in their own documents, checks run automatically before every build, and decisions are logged with a date and a rollback command.

The tooling has improved along the way too: when something turned out badly, the rule or skill that let it through was tightened.

The way of working is written down so that a team can pick it up and use it.