Keep this open beside the chat window and run through it before you hand off a big, ambiguous task. Every box below is a decision you make now, while it's cheap, instead of a question the agent has to interrupt you with later. When they're all ticked, you can type "go" and close the laptop. This is the pass I run before I let anything work unsupervised.
1. Mission — the goal as an outcome
- [ ] Write the goal as the result you want to hold, not a list of steps to perform.
- [ ] Keep it to a sentence or two: the outcome, and one line on why it would work.
- [ ] Grant real creative latitude — one line like "within the limits below you have total freedom; I want your best work, not your safest." A short leash buys cautious, forgettable output.
Goal (outcome): ______________________________________________
2. Guardrails — what makes leaving the room safe
- [ ] Spend nothing new — work only with what you already have.
- [ ] Publish nothing — build everything locally, so a mistake stays on the machine instead of going out under my name.
- [ ] Invent nothing — every fact and figure checked against a real source, no plausible-sounding filler.
- [ ] Least access — only the keys the project already has wired in; nothing beyond that.
My hard limits: ____________________________________________
3. Phases — the shape of the work
- [ ] Lay out the work as an ordered arc, first step to last.
- [ ] Add the line that keeps it from becoming a cage: this is a floor, not a ceiling — if you see a step worth adding, add it.
Phases (in order): _________________________________________
4. Deliverables — the artifacts I expect to hold
- [ ] Name each concrete artifact, not "a writeup about it" (the site, the working product, the plan, the research).
- [ ] Require one recap document that links to all of them, so the pieces arrive tied together.
I should receive: __________________________________________
5. Definition of "done" — the part that does the heavy lifting
- [ ] Write the bar as something a stranger could check without asking me.
- [ ] Make each point a test, not an opinion: open X, run Y, confirm Z. Without this line the agent quits half-built or polishes one corner forever.
Done = when a stranger can: ________________________________
6. The "never ask" rule — the line that lets you close the laptop
- [ ] Make every call yourself, and write down why for each one.
- [ ] Don't come back until the definition of done is met.
- [ ] Earned it? Only hand this over once the guardrails make every unsupervised call safe and the done line is genuinely checkable. Without those two, "never ask me anything" gets you confident nonsense, fast.
7. Orchestration (optional) — how to work, as a floor not a ceiling
Patterns, not step-by-step instructions. Hand over the ones that fit and let the agent add its own.
- [ ] Fan out parallel researchers across different sources and angles, so one blind spot doesn't quietly sink the picture.
- [ ] Run an idea tournament: competing directions pitched by independent agents, a judge panel scoring each against plain criteria.
- [ ] Put a skeptic on every claim that matters, with one job — tear it down. What survives, you can lean on.
- [ ] Close each phase with a completeness critic before the work is allowed to move on.
- [ ] Let the capable model manage — plan, delegate the building to cheaper models, review, send back — instead of grinding through everything itself.
8. Final gate before "go"
- [ ] Is the mission an outcome I want, not a task list?
- [ ] Can a stranger check the done line without me in the room?
- [ ] Do the guardrails make walking away actually safe?
- [ ] Did I stress-test the idea itself before pointing an agent at it? A capable agent builds a weak idea beautifully — that's the trap.