How I work

Agents write a lot of my code. I wrote the rules they work under, I review what they ship, and nothing goes out under my name that I can't explain.

Eight rules, eight gates

Hover a gate for its rule. Click a change to stop it yourself. Tap a gate for its rule. Tap a change to stop it yourself.

1 2 3 4 5 6 7 8
  1. 1

    The board goes up before the first edit.

    I run several agents at once, and the board is how I see what each one is doing.

  2. 2

    Never guess a fact a lookup could settle.

    Every claim rests on something read in that session.

  3. 3

    A comment says why, or it doesn't exist.

    A comment that narrates the code is the first thing to go stale.

  4. 4

    Files change through file tools, never the shell.

    So every change is one I can read before it lands.

  5. 5

    Nothing is created under my name unless I asked for it.

    No ticket, merge request or pushed branch shows up as a side effect.

  6. 6

    Fix the cause, not the symptom.

    Cleaning up the damage is step one. The fix is whatever made it possible.

  7. 7

    Whether it looks right is my call.

    Agents prove the logic with types, builds and tests, and I look at what people will see.

  8. 8

    A rule I keep breaking is a missing hook.

    A rule that matters gets a script that stops the agent, not another sentence.

From the ask to shipped

The agent reads what I typed, what I meant and what it forces. It puts the board up, builds in small steps, and proves each one with a check that could disagree. I review it before anything leaves.

  1. Read what I typed, what I meant, what it forces
  2. Board #4 [AI] Crossfade the sequence frames
  3. Build fix(scroll): crossfade the two frames around the scroll position
  4. Prove check.ts: 0 failures
  5. Review
  6. Ship merged into main