真正适用于使用 AI 智能体的独立创始人的委派简报

If you are a solo founder (or a very small team) shipping more than one product, you already know the trap: AI chat feels productive, then you open the folder and nothing usable is there. Fluency is not delivery. The fix is not a smarter model first — it is a sharper brief.

This excerpt is adapted from The Solo Founder's AI Agent Playbook (practical operating notes for one person running agents without losing the plot, the budget, or honesty about what actually shipped).

Agents are junior specialists, not autopilot

A solo founder with agents is not "automated." You are a manager of unreliable junior specialists who work at machine speed. Treat them that way.

Name the roles even if one agent wears several hats:

• Operator (you): sets goals, approves spend, ships final calls, owns customer promises.
• Researcher: gathers sources, compares options, flags uncertainty.
• Builder: drafts code, docs, listings, emails — never silently publishes.
• Editor / QA: checks facts, tone, and "would I stake my name on this?"
• Ops runner: schedules, files, status reports, checklist execution.

Separate agents win when context windows fill up or when you need a second opinion that did not write the draft. Collapse roles only when the work is tiny.

Non-negotiables that belong in every brief:

• Agents propose; you dispose — especially for money, public posts, and legal claims.
• Every lasting artifact lives outside the chat (repo, drive, product folder).
• Secrets never go into prompts you would not paste into a shared doc.
• If a metric cannot be shown from a primary source, it is not reported as fact.

The brief is a one-page contract

Vague prompts produce vague work. Fill these fields every time:

If you cannot write the brief, you are not ready to delegate the work. That sentence alone saves more money than any prompt pack.

Example brief (product PDF)

OUTCOME: 4-page Part 107 cheat sheet PDF + 1280x720 cover + listing.md
CONTEXT: /workspace/products/_shared brand; facts only from bank.json + lessons
DONE: accurate 14 CFR cites; no lorem; page-1 PNG looks clean
MUST NOT: spend money; invent testimonials; claim FAA affiliation
SOURCES: eCFR, existing lesson recaps, question bank
DEADLINE: same day; report page count + file paths
HANDOFF: Operator reviews page-1 render before listing goes live

Write briefs you would give a sharp contractor on day one. Agents amplify whatever clarity (or fog) you hand them.

Route work so it does not bounce forever

Without routing rules, agents duplicate work or wait on each other forever. A simple router is enough:

1. New idea → Researcher (1-page memo: demand, competition, build cost, kill criteria).
2. Approved idea → Builder (artifacts only).
3. Draft ready → Editor/QA (checklist, not vibes).
4. Passed QA → Operator ships or schedules.
5. Live product → Ops runner (daily status, support drafts, promo calendar).

Conflict rules that prevent silent disasters:

• One writer per artifact tree. Two agents editing the same PDF is how you get silent overwrites.
• Questions about money, legal claims, or public identity escalate to you — never "just send it."
• If an agent is blocked more than ~30 minutes on a credential or approval, it stops and reports the block instead of inventing a workaround.

Handoff format (keep it short)

STATUS: Draft PDF v1 ready
PATHS: /workspace/products/foo/foo.pdf, cover.png, listing.md
UNVERIFIED: one regulatory cite marked [CHECK]
NEED: Operator visual check of page 1; approve price $9

No novel-length chat dumps. Paths beat vibes.

Review loops that assume errors

Agents are fluent. Fluency is not truth. Run three passes:

• Pass A — Structure: Does the artifact match the brief? Missing sections? Wrong audience?
• Pass B — Facts: Every number, cite, price, and claim maps to a source. Unsourced → cut or mark.
• Pass C — Taste / risk: Would you put your name on this publicly? Any secret leaked?

Red flags worth training yourself to spot: round numbers with no source, testimonials that were never collected, stale "as of 2024" language in a 2026 doc, tools described as installed that are not on the machine, and spend recommendations that ignore your hard budget cap.

Make the Editor argue against the Builder. Agreement is suspicious; productive disagreement is useful.

Start today

1. Pick one real task you would otherwise do yourself.
2. Fill the brief table above — especially Must not and Done looks like.
3. Route Builder → Editor → you.
4. Demand file paths in the handoff.
5. Cut anything unsourced before it reaches a customer.

That is the operating core. Templates and checklists make it reusable; the brief is what makes it real.

Want the full operating system — brief templates, routing rules, review loops, budget discipline, launch checklists, and the copy-paste prompt appendix?

Solo Founder's AI Agent Playbook — $19 digital PDF on Gumroad.