Chapter 02 · Named Missions

Field Procedures

Sixteen repeatable workflows in four mission phases — prepare, build, verify, scale. Each is a bounded mission: the move, the steps, the moment to reach for it. Procedures are durable; the specific tools inside them live (stamped) in the Command Console.

Eggo the launch director dispatching courier drones along separate neon flight paths while a returned drone presents its proof-chip
every mission returns a proof-chip

A workflow you can't name is a workflow you can't repeat.— why these have call signs

Ⅰ

Prepare

Eggo the mission briefer tapping a pointer on a holographic route diagram with waypoints lighting up and drones ready to launch

Know the plan before the launch. Every expensive agent failure in fifteen months of workshops traces back to the same origin: execution started before the mission was defined. These four procedures are how a mission gets defined.

P·01 — Scaffold → Outline → Instruct → Execute → Monitor

Stage the work; never freestyle a feature.

Start from an official scaffold and inspect current reality. Produce an outline; turn it into explicit instructions. Execute in bounded increments; monitor and correct. Use when: building anything non-trivial or unfamiliar. This staging survived every tool era unchanged.

P·02 — Forced Context

Hand-pick the evidence; never dump the repo.

Search for a precise concept. Open the results together, remove the noise, keep paths and line refs. Supply only the curated evidence. Use when: a change spans files but the relevant surface can be found mechanically.

P·04 — AI Interview / Rule of Threes

Make the agent ask before you make it act.

Ask for three questions, critiques, or alternatives. Answer by voice or text. Generate variants; select with explicit criteria. Use when: requirements are fuzzy or taste matters.

P·05 — Gather → Document → Build

Real examples first; synthesis second; code last.

Find real examples via source search and docs. Synthesize a best-practices artifact. Implement from that artifact; verify locally. Use when: the library or API is unfamiliar.

Ⅱ

Build

Eggo the machinist feeding a glowing speech-bubble orb into a fabrication machine that outputs a sparkling new multi-tool

Ship in bounded, reversible moves. The build phase is where enthusiasm outruns discipline — so every one of these procedures has a built-in brake.

P·07 — Hypothesis Mode

One hypothesis, one experiment, one verdict.

State the hypothesis. Isolate the experiment; make the smallest change; observe. Revert or refine — and name the pattern only after success. Use when: tool behavior, architecture, or automation value is uncertain.

P·08 — Plain-English Tool Factory

Prompt in, tool out.

Make a prompt work in print mode. Constrain its tools; wrap it in a function; expose parameters. Structured output only when a consumer needs it. Use when: the same judgment task keeps recurring.

P·09 — Clean-Session Task Handoff

Pass the capsule, not the desk.

Research broadly. Save the chosen plan and acceptance criteria to a task artifact. Start a fresh agent; implement only from the artifact; review against it. Use when: research context is huge or littered with rejected paths.

P·12 — Diagram-Driven Development

If it crosses a boundary, draw it first.

Describe front-stage behavior and back-stage implementation in Mermaid. Store diagrams with the project; inject them into planning; update after structural changes. Use when: features cross architectural boundaries. (Diagram-Driven, not Domain-Driven.)

P·06 — Video → Instructions → Code → Browser Proof

Show the behavior; don't describe it.

Record the behavior. A multimodal model extracts structured requirements; the coding agent implements; verify in the browser. Use when: the problem is visual or interactive. Honest label: least battle-tested procedure here — promising, not proven.

Ⅲ

Verify

Eggo the white-gloved inspector scanning a freshly assembled robot's seam, the beam catching one amber flaw the builder missed

Prove it, or it didn't happen. Every procedure in this phase exists because an agent once said "done" and it wasn't. Fresh eyes, saved receipts, independent contexts — the theme is never let the builder grade its own work.

P·10 — Implement → Independent Review

The builder never grades the build.

Implement; commit. A separate context inspects the diff or latest commit cold. Address findings; rerun verification. Use when: any bounded change can be reviewed independently — which is most of them.

P·13 — Multi-Lens / Council Review

Different lenses, not more of the same lens.

Generate distinct review lenses; run them independently; compare overlap and disagreement; apply human taste; preserve the decision. Use when: architecture, product, accessibility, security, or design deserves real scrutiny. And remember: a missing verdict is not agreement.

P·14 — Browser Explore → Deterministic Script

Agents explore; scripts repeat.

Let the agent explore interactively. After several successful repetitions, encode the stable path in Playwright. Keep agent judgment only where variability remains. Use when: a browser workflow is recurring and stable.

Eggo the crash investigator lifting a glowing memory chip from a charred drone with tweezers while a pristine replacement drone waits

P·03 — Session Post-Mortem

Don't keep flying the crashed drone.

When a session drifts, repeats itself, or hauls too much debris: stop flying it. Export the session or read its history. Ask a fresh (or stronger) model to diagnose the failure. Save the useful diagnosis — that's the chip worth rescuing — and restart from a clean context.

The wreck is evidence. The next flight is fresh. Both facts matter equally.

Ⅳ

Scale

Multiply without losing the thread. Scale is the phase with the most spectacular failures — because it's the only phase where the bottleneck is you. Three procedures, all shaped by the same constraint: human attention doesn't scale with token spend.

P·11 — Automation Packaging Ladder

Climb only as the value repeats.

Prove the prompt interactively; save it. Wrap it in a shell function. Move complex orchestration into a normal program. Promote invariants to hooks or CI. Use when: a workflow has demonstrated repeated value — never speculatively.

P·15 — Manager / Worker + Specialists

One coordinator holds the goal; workers hold fragments.

Keep the goal and decision history in a coordinator. Dispatch narrow tasks with bounded inputs. Workers return artifacts and receipts; integrate centrally. Use when: work divides without overlapping ownership.

P·16 — Browser Bridge / Controlled Variation

Consult wide, execute local, prove everything.

Preview a bounded context bundle; send it through the browser-model route for a broad plan or critique. Execute locally; save verification receipts. For alternatives: one meaningful dimension per lane, shared proof contract. Use when: a high-context model adds leverage but execution and proof must stay local.

Every mission returns a proof-chip — or it didn't return.— the drone bay rule