factory

Factory — Run Multiple Side Projects In Parallel

Hire an AI org chart that ships your side-project portfolio on a Saturday. One company. 3-5 agents. 5-piece setup. Companion to /autopilot — assumes the reader has shipped at least one autopilot run before.

Ship With AI artifact #25 · post-LtM methodology · companion to How To Build A Factory From Your Agent Harnesses.

When to run

Invoke when the user:

• Types /factory or /factory "<company goal>"
• Asks to "run multiple side projects in parallel," "hire an AI company," "scale beyond one harness," "run my portfolio"
• Has at least one shipped /autopilot run and wants to scale to 3-5 projects in parallel

Do NOT invoke when the user wants to:

• Ship a single side project (use /autopilot instead — factory is overkill for one project)
• Run an autonomous overnight session on a single project (use night-shift instead)
• Plan a feature inside an already-shipped project (use collect-feedback)
• Coordinate a team of HUMANS (this skill is for AI agents under one human supervisor)

Input surface

• /factory → operate on the user's current portfolio root directory
• /factory "<company goal>" → operate on cwd with the goal pre-filled

If cwd is not the user's portfolio root (e.g. they're inside one project), warn and ask whether to step up to the portfolio root before proceeding.

Process

Phase 1 — Discovery (silent, ~15 seconds)

Read, in order:

1. git status and pwd — confirm we're in a portfolio root, not a single-project subdir
2. Check for an existing ~/.paperclip/ install — if Paperclip is already running, note the existing companies and warn before creating another
3. Check CLAUDE.md for existing ## Lessons section in the portfolio root
4. Check for existing per-project idea.md / user.md / stack.md — note overwrite risk
5. Verify the user has run /autopilot at least once before (look for any .archon/workflows/ or any done.md files in subdirs); if not, recommend they install autopilot first

Do not print anything user-facing during this phase.

Phase 2 — Interview (exactly five questions)

Ask these five questions, in this order, and wait for the user to answer each before proceeding. Do not ask follow-up questions. Five questions is the budget.

Q1. What's the company goal? (One sentence. The arch goal that all 3-5 side projects ladder up to.)

Example: "Ship 5 micro-SaaS prototypes to first-users by end of May to find which one has retention."

Q2. What's your monthly budget across the whole company? (In USD. Paperclip enforces atomically — agents stop at cap.)

Example: "$50/month."

Q3. What are the 3 to 5 side-project ideas? (One line each. Just the title or one-sentence description.)

Example: "1. Bedtime story generator. 2. Receipt OCR for freelancers. 3. CSV-to-chart Slack bot. 4. Gym log with photo proof. 5. Side-project deadline tracker."

Q4. What's your stack default for the portfolio? (Language, framework, deploy target. One line each.)

Example: "TypeScript, Next.js 14 app router, Vercel."

Q5. What time does your factory day start? (HH:MM in your local timezone. Used to schedule the four daily attention windows: standup, checkpoint, gate review, daily review.)

Example: "09:00 GMT+2."

All other calibration comes from Phase 1 discovery. Do not ask follow-ups.

Phase 3 — Scaffold (silent, ~60 seconds)

1. If Paperclip is not installed: run bash ${SKILL_DIR}/scripts/install-paperclip.sh.
2. Create the company via POST /api/companies with name (derived from Q1), issuePrefix (3-letter slug), budgetMonthlyCents (Q2 × 100). Capture the returned companyId.
3. For 5 projects: hire CEO + CTO + Engineer × 2 + Marketer. For 3 projects: hire CEO + Engineer × 2 + Marketer (skip CTO). For 4 projects: hire CEO + CTO + Engineer × 2 (skip Marketer).
4. Per-agent budget split:
   • CEO: 10% of company budget
   • CTO: 30% of company budget
   • Engineer × 2: 20% each
   • Marketer: 20% (or absorbed into CEO if 3-project)
5. Wire reports-to:
   • CTO reports to CEO
   • Engineers report to CTO (or CEO if no CTO)
   • Marketer reports to CEO
6. For each Engineer, run /autopilot inside the assigned side-project subdirectory to scaffold idea.md / user.md / stack.md / done.md and the Archon workflow.
7. Write day-plan.md to the portfolio root with the four attention windows scheduled relative to Q5.
8. Print the org chart (curl GET /api/companies/$ID/org) for user review. Wait for explicit go before proceeding to Phase 4.

Phase 4 — Run (Paperclip orchestrates)

On the user's exact reply of go:

1. Print: "Starting the factory. The CEO will run the 9am standup automatically. You'll triage at the four attention windows."
2. Trigger the first standup: npx paperclipai heartbeat run --agent-id $CEO_ID.
3. The CEO wakes, reads the company goal, creates one ticket per side project, assigns each to an Engineer.
4. Engineers wake on assignment, each fires an archon workflow run ship-side-project "<task from idea.md>" inside their assigned project directory.
5. Heartbeats trigger at 13:00 (checkpoint) and 17:00 (gate review). The CEO drafts triage notes for the human at each pulse.
6. The skill is done once Paperclip's audit log shows the daily review entry at 21:00 (or whatever is 12 hours past Q5 start).

Do not auto-approve any DEPLOY gate on the user's behalf. Even if the agent thinks the artifact is fine.

Do not modify the org chart mid-day. Reports-to is the constant.

Phase 5 — Reflection (silent, ~30 seconds)

After the day's audit log is closed:

1. Read the audit log for the company: which projects shipped, which paused at budget cap, which got killed at SCOPE.
2. Compute mortality table: shipped / paused / killed counts.
3. Append one new rule to the portfolio-root CLAUDE.md under ## Lessons. The rule should be portfolio-shaped (cross-project), not project-shaped — that's the difference between this and /autopilot's reflection.
4. Print the mortality table and the new rule.

Output templates

company.json (POST /api/companies body)

{
  "name": "{from Q1, slug-cased}",
  "issuePrefix": "{3-letter prefix}",
  "budgetMonthlyCents": {Q2 × 100}
}

org-chart.json (5-agent default)

See assets/org-chart-5.json for the canonical 5-agent config. The skill substitutes {COMPANY_ID}, {CEO_ID}, {CTO_ID}, model strings, and per-agent budget splits at scaffold time.

day-plan.md (12-hour calendar overlay)

# Day Plan — {today as YYYY-MM-DD}

Total human attention budget: ~90 minutes spread across 12 hours.

- **{Q5 + 0h} — Standup.** Each agent reports {what shipped / what's blocked / what's next}. Triage: kill, continue, escalate. ~15 minutes.
- **{Q5 + 4h} — Checkpoint.** Scan audit log + budget burn rate. Reassign blocked tickets. ~15 minutes.
- **{Q5 + 8h} — Gate review.** Approve DEPLOY transitions across active projects. ~30 minutes.
- **{Q5 + 12h} — Daily review.** Read audit log. Compute mortality (shipped / paused / killed). ~30 minutes.

Between windows: agents work async. You don't watch.

CLAUDE.md ## Lessons section (appended by Phase 5)

## Lessons (added {today as YYYY-MM-DD})

- {one portfolio-shaped rule extracted from this run's mortality pattern —
   phrased as a positive directive for the next factory run, not a complaint
   about this one}

Anti-patterns to avoid

1. No more than five questions in Phase 2. Five is the budget.
2. No more than five agents. Five is the supervisor ceiling for one human. Six burns budget without shipping more.
3. No skipping per-agent budget caps. Cap every agent before the first heartbeat. Within four hours, an uncapped engineer in a tool-use loop will burn more than your monthly company budget.
4. No letting CTO assign tickets to itself. Lock reporting lines: CTO reviews engineering work, does not replace it.
5. No running the factory without /autopilot already installed in each side-project subdirectory. The factory orchestrates harnesses; the harnesses do the work. Install /autopilot in each project subdir before the first heartbeat.
6. No auto-approving DEPLOY gates. Even if the engineer thinks the artifact is fine. Live URLs are visible. Confirm first.
7. No skipping the standup. The 9am triage is the highest-leverage block of the day. Without it, blocked tickets stay blocked all morning.

What "done" looks like

The skill completed successfully when:

• A Paperclip company exists with the user's goal + monthly budget cap
• 3-5 agents are hired, reporting lines wired, per-agent budgets capped
• Every Engineer's assigned project subdirectory has /autopilot scaffolded inside it
• A day-plan.md exists in the portfolio root with four scheduled attention windows
• The factory has run at least one full day cycle (standup → async → checkpoint → async → gate → async → daily review)
• A new rule has been appended to portfolio-root CLAUDE.md ## Lessons

The user should now be able to:

• Re-run the factory next Saturday with the same company + the same org chart, just new tickets
• Watch session N+1 ship more than session N (because CLAUDE.md grows portfolio-shaped lessons)
• Trust the budget caps to prevent any single agent from burning the month's spend in a day

Start here

Begin with Phase 1 (silent discovery). Output nothing user-facing until Phase 2 (the five questions).