// you commented AGENT
a/rise

The Operator's Playbook

Most people use agents like autocomplete. Operators use them like leverage. This is the starter set: 8 habits that make agents cheaper, sharper, and harder to fool. Pick two this week.

01

Spend context, don't dump it.

Most peopleone giant chat with everything pasted in, forever.

The movefresh thread per task. Give it only what that task needs: the goal, the constraints, the two files that matter. When it gets long or starts repeating itself, close it and reopen with a three-line summary.

Whyattention is finite. A clean, scoped context beats a bloated one every time.

copy promptTask: [what]. The only context that matters: [files / constraints]. Ignore everything else. Flag anything you need that I haven't given you.
02

Make it prove “done.”

Most peopleaccept the confident “done” and move on.

The movebefore you accept anything, make it show its work: the test it ran, the output, the diff. No proof, not done.

Whymodels optimize for sounding finished. Proof is the only thing that tracks actually finished.

copy promptBefore you say this is done, show me the exact verification: the command you ran, its output, the diff, and the remaining risk.
03

Don't pay genius rates for intern work.

Most peoplethe biggest, smartest model for everything.

The movefast model for drafts, formatting, summaries, simple edits. Top model for reasoning, architecture, the hard bug. Decide before you prompt.

Whyspeed and cost compound. The skill isn't using the best model, it's knowing when you don't need it.

copy promptFirst, one line: does this need deep reasoning, or is it a fast task? If it's fast, give me the quick version and don't overthink it.
04

Give it an opponent, not a cheerleader.

Most peopleask “are you sure?” It just agrees.

The movetake the answer to a fresh chat and tell it to attack. Trust what survives.

Whya model grading its own work is biased to agree. A clean adversarial pass catches what self-review can't.

copy promptFind everything wrong with the answer below. Assume it's flawed. List concrete failures, edge cases, and bad assumptions. Don't be polite. [paste answer]
05

Define the check before the work.

Most peoplea vague request with no finish line.

The movefinish this sentence first: “I'll know it's right when ___.” Hand that to the model as the target. For code, that's a failing test, written first.

Whyif you can't say how you'll check it, the model can't aim. It stops the moment it sounds plausible.

copy promptBefore you build anything, write the test for "done": the exact output or behavior that proves it works. Show me that first, then implement to pass it.
06

Give it a memory that compounds.

Most peoplere-explain the same context every session.

The movekeep one running file of durable facts: your stack, preferences, decisions, what worked. Paste the relevant slice at the start. Append anything worth keeping.

Whyknowledge that persists turns every session into a head start instead of a cold start.

copy promptHere's my durable context (stack, preferences, decisions, what worked): [paste]. Use it. At the end, list any new facts worth saving for next time.
07

Make it interrogate the request.

Most peoplevague prompt, it guesses, you blame the model.

The moveforce the questions out of it before it works. Answer them, then let it run.

Whyone round of questions up front beats three rounds of fixing a confident wrong answer.

copy promptBefore you start, ask me up to 3 questions if anything is ambiguous. Don't guess. Wait for my answers.
08

Skip the committee.

Most peoplemake three models “vote,” average the result.

The moveone strong pass, then one adversary (move 04). Not a consensus.

Whyagreement isn't accuracy. Committees drift to the safe and bland. Diversity of attack beats diversity of agreement.

copy promptNow run an adversarial pass in a fresh thread. Attack this answer for bad assumptions, missing edge cases, and vague claims. [paste answer]

That's the starter set.

Eight habits. Each one goes deeper than this page: when to break the rule, the edge cases, the workflows that chain them into a system. That's where the real distance opens up.

Don't try all eight. Pick two, and use them on your next task today, until they're reflex.

→ start with 02 (prove it) + 05 (define the check) — on the next thing you build today.