Mark Levy/Writing/The Word Warp Drive

An 8-bit observation lounge: officers around a curved conference table beneath tall windows full of stars, a wall display showing a checklist with amber checks and one red cross, and four lights in a row on the bulkhead.

[ Tier 2 · Deck 2 of 2 ] Command: getting the most out of the crew

The Word Warp Drive

Mission orders, sensor readings, course corrections, the first officer, and command authorization: the craft of the captain's chair.

New here? Tier 2 assumes Tier 1, and this deck assumes Deck 3 first.

Q
You typeShip the Basic tier
Same loop as Deck 3. This is Tier 2 command: what the captain does around it.

Deck 3 gave the core hands. Deck 4 closes Tier 2 with command habits. Three facts drive everything: the word warp core predicts plausible continuations (Deck 1), verifies nothing itself (Deck 1, Section 6), and carries no memory between sessions (Deck 3, Section 6).

Command turns those constraints into shipped work: define done before engage, trust readings over reports, correct in short hops, request fresh review, and hold the human boundary. Five sections, five demos, one Cardassian at the bottom.

QRewatch any episode: Picard rarely touches a console. He asks, orders, then scrutinizes reports. Command is mostly before and after "Make it so."

Section 1 of 5Mission Orders

Say what done means before you say engage

The core targets what you write. Anything omitted gets plausibly guessed, which is not always what you meant. "Improve the login page" has many plausible completions, so define done first as checkable criteria.

Put that list in a file, not chat. Chat compacts and sessions end (Deck 3, Section 6); files persist and get reread. Keep criteria short: long briefs dilute attention (Deck 1, Section 4).

The orders desk: what you wrote, what got guessedhand-built demo, canned report
The orderImprove the login page.
Criteria in the plan file
Mission report
Captain's log

Unspecified details get plausible guesses. Write done as short, checkable criteria in a file before engage.

Back on Earth Spec, acceptance criteria, definition of done, PRD, or checklist ticket are all mission orders. Plan mode enforces this step before edits. Spec-driven development keeps one file as shared target for fresh sessions and away teams.
Q"Tea. Earl Grey. Hot." Minimal words, no ambiguity. "Make it better" invites plausible beverages.

Section 2 of 5Sensor Readings

A report is text; a reading is evidence

"All 14 tests pass" is a plausible ending shape. The core can write it whether tests ran or not (Deck 1, Section 6). Not deception—just no direct world access. A report is text; a reading is tool evidence: test output, screenshot, diff, successful URL response.

Separate them deliberately. Request readings for each claim, and automate readings via hooks that run tests and paste results back (Deck 3, Section 5).

The mission report: five claims, five readingshand-built demo, canned evidence
The report, as written by the crewFixed the parser, all 14 tests pass, verified on a phone, no API changes, deployed to the preview channel.
    Five sentences, each the plausible shape of a finished job. Nothing has been checked.
    Captain's log

    A report is text; a reading is evidence. Claims without readings are holodeck output. Wire readings into the loop automatically.

    Back on Earth CI output, test logs, target-width screenshots, the diff, and curl checks are readings. Hooks automate collection. LLM-as-judge is useful but still bettor-grading-bettor. "Show evidence" is the shortest good command here.
    QThe Drumhead. Admiral Satie's tribunal ran for two days on confident assertion and ended the moment somebody asked for a reading. Picard's line about it was that the chain is forged one link at a time. The first link was a claim nobody checked.

    Section 3 of 5Course Corrections

    Small hops, frequent fixes

    Each crew step is a bet (Deck 1). Bets stacked on bets compound drift: one wrong assumption can shape dozens of later turns. The loop rarely notices because each step stayed plausible relative to the previous one.

    So measure often: small hops, quick readings, frequent commits. Then bad steps are reverted, not patched. Autonomy is a cost dial: steps between checks.

    The drift plot: forty steps, one coursehand-built demo, seeded noise
    The dashed line is the course you plotted. Pick an interval and run.
    Captain's log

    Unchecked errors compound. Use short hops, readings after each, and checkpoints. Revert beats repair when history exists.

    Back on Earth Small PRs, early commits, red-green-refactor TDD, and git reverts are this pattern. Agent autonomy is the same dial: higher for cheap reversible work, lower for costly steps.
    QContagion spread one subsystem at a time. Geordi fixed it by reload, not patching every symptom. Revert worked because a clean copy existed.

    Section 4 of 5The First Officer

    Fresh eyes don't share your blind spots

    By turn forty, context has absorbed your assumptions. Ask that same context to review itself and you often get agreement with familiar reasoning, not a real second opinion.

    A first officer review uses empty context (Deck 3, Section 4): diff, orders, relevant doc, nothing else. Different vendors share fewer blind spots. Delegate independent, fully briefed, verifiable tasks—and read the report critically.

    Two reviewers, one diffhand-built demo, canned reviews
    // ledger.ts: the ledger stores minor units - const cents = item.price; + const cents = Math.round(item.price * 100); ledger.post(order.id, cents);
    Store API doc, line 14: price is an integer in minor units (cents).
    Turn 12 of the bridge conversation says: "the store API returns dollars." It doesn't. Pick a reviewer.
    Captain's log

    Fresh eyes reduce shared blind spots. Use empty-context review, ideally cross-vendor. Delegate only work that is independent, briefed, and verifiable.

    Back on Earth Fresh-context review subagents, cross-model audits, diff-only PR bots, and adversarial review are first-officer patterns. Mirror warning sign: review cites your own prior chat as evidence.
    QRiker's value is disagreement in the ready room. A first officer who only says "Aye" is just a decorated mirror.

    Section 5 of 5Command Authorization

    Automate to the boundary, then use your voice

    Sort actions by two tests: is it reversible, and who does it affect? Reversible contained work can run automatically; irreversible or outward-facing work needs human approval; a few actions should never be delegated.

    Move the boundary outward only where readings justify it, then hold the line. Preview environments help by allowing end-to-end checks before irreversible release.

    The authorization board: eight actions, three shelveshand-built demo, opinionated defaults
      Assign each action to a shelf: runs on its own, asks you first, or never delegated.
      Captain's log

      Automate to the boundary, then use your voice. Reversible contained work runs itself; irreversible outward work waits for you. Previews move boundaries, not responsibility.

      Back on Earth Permission modes and allowlists draw the boundary. PR previews, staging, and dry runs move it outward. Human-in-the-loop enforces it. Ask every action's blast radius: who is affected, and can it be undone?
      QAuto-destruct needs two voices and a countdown. Nobody automated that, yet the ship automated plenty else. That is the line.

      HazardsKnown hazards on this deck

      Four ways command goes wrong

      Known hazards on this deck
      • HAZARD 1Aye-Captain drift. Context that agrees with itself keeps agreeing. Sycophancy is often continuation pressure, not personality.
      • HAZARD 2Plausible done. Finished-sounding reports without readings. Ask what tools actually returned.
      • HAZARD 3Scope creep. Crew improves adjacent code nobody ordered. Done criteria must include not-in-scope.
      • HAZARD 4Endless loops. Impossible checks cause repeated retries. Use turn budgets and explicit stop-and-ask rules.
      Captain's log

      These hazards are normal continuation behavior. Agreement, tidy-ups, and persistence are plausible outputs; command predefines what gets checked.

      ManifestThe command vocabulary

      What's in the cargo bay

      Vocabulary for working well with agents, with deck addresses:

      Manifest: 14 items
      • spec · acceptance criteriaShort checkable done-list written before editing.
      • definition of doneThe finish line, proven by readings.
      • plan filePersistent orders on disk; chat forgets, file remains.
      • verification · evidenceTool readings from the world, not report prose.
      • TDDWrite failing check first, then satisfy it.
      • checkpoint · revertCommit verified hops so bad steps are quickly undone.
      • the autonomy dialSteps between checks; set by risk and reversibility.
      • fresh-context reviewFirst-officer review from empty context.
      • cross-model reviewReview from another vendor to reduce shared blind spots.
      • blast radius · reversibilityWho gets affected, and can you undo it?
      • permission modeBoundary policy: run, ask, or refuse.
      • preview deploy · dry runSafe realism that moves boundaries outward.
      • sycophancy · scope creepAgreement loops and unasked-for improvements.
      • turn budgetLoop cap that forces stop-and-ask over endless retries.
      Captain's log

      No item here changed the core. These are captain habits with text, files, and boundaries. Same crew, better chairwork.

      TribunalGul Madred

      Tribunal: five charges, three shield levels

      Q
      Order. Prosecutor: Gul Madred, who insists there are five lights and has a very confident report. Count for yourself; that's this deck. Five charges. Right answers dismiss one; wrong answers drop shields and point to review. Lose all shields and I snap you to the door.
      The courtroomthere are four lights

      Deck clearedWhat this buys you

      Four commendations

      Four rules fall out of these five sections, and teams keep relearning them the expensive way:

      • Write done before engage. Short, checkable criteria in a file before first edit.
      • Readings over reports. Reports are text; readings are evidence. Automate readings into the loop.
      • Short hops, checkpoints. Check often, commit verified steps, and revert faster than you repair.
      • Automate to the boundary and hold it. Reversible contained work runs itself; irreversible outward work needs your voice.

      That's the chair: same core from Deck 1, same crew from Deck 3, better command. Write destination, read instruments, correct in small hops, seek fresh review, hold the boundary—and the crew ships.

      Turbolift

      Tier 3 begins on Deck 5: Temporal Mechanics. How far and how fast this goes: curves beneath hype, smaller ships, robots, the word "general," and what to do next.

      Turbolift to Tier 3 (Deck 5) Back to the ship