A resumable orchestration system for long-running Claude workflows

A developer released Incubator, a resumable orchestration system for Claude that enforces a seven-phase sequence—interview, research, report, hire, board meeting, data room, and dashboard—to prevent premature validation in long-running AI workflows. The system acts as a founding CEO, generating specialist teams and maintaining session continuity via an index file.

A Claude skill that acts as the founding CEO for a new project — interviewing you into a testable hypothesis, researching the market, writing a findings report that will tell you to pivot if the evidence says so, and generating a sequenced specialist team: persona briefs and installable skill files derived from the research, one per role. Built for Claude Cowork https://claude.ai , Claude Code, and claude.ai. Agentic coding makes building feel free. That's the trap — you ship a prototype in an afternoon and mistake it for validation. Sense-making falls behind building. Incubator enforces the order. It won't let you hire a team before the research exists, or build before the idea is a falsifiable hypothesis. The CEO is required to tell you when the evidence says stop. And when you come back a week later, it picks up exactly where the company left off. ⭐ If this is useful, star the repo — it helps others find it. Invoke project-ceo at the start of any project. Claude becomes the founding CEO for your specific vertical — with real industry expertise, not generic advice. It runs a seven-phase sequence: | Phase | Mode | What happens | |---|---|---| 1. Interview | Interactive | CEO interrogates the idea one question at a time — exits only when you agree on a testable hypothesis. Captures your time budget and VC intent. | 2. Research | Fully autonomous · hard cap 20 searches | Maps competitors by tier, reverse-engineers 2–4 winners' build sequences, digs the graveyard for causes of death, runs a mandatory devil's-advocate pass | 3. Report | Autonomous | Research report: 9 panels for startups, 5 for side projects/MVPs. Thesis → Analogues → Graveyard → Heresy → The Call — or the full arc to Trends and Org. | 4. Hire the Team | Interactive | Derives a 4–7 agent roster from your vertical, sequenced by what de-risks fastest. Quality-gates every generated skill — no flavor without substance. | 5. Board Meeting | Ongoing | Per-agent 🟢/🟡/🔴 health scores, roster changes, riskiest current assumption, next action. Runs at every milestone or returning session. | 6. Data Room | Conditional VC track | Converts research into a 10-slide pitch deck weighted on Graveyard and Playbook — why others failed, why this sequence wins. | 7. HQ Dashboard | Automatic | Generates a zero-dependency dark-mode hq.html — your command centre. Charter, roster, and hiring plan in one browser tab. | The CEO doesn't disappear after founding. It writes an INCUBATOR.md index at your workspace root. Every time you return — "CEO check in", "let's work on the project" — it re-reads the roster, checks team health, and picks up exactly where the company left off. The examples/airbnb/ /afsalali1238/Incubator/blob/main/examples/airbnb folder shows real output from a full run on the founding of Airbnb: | File | What it shows | |---|---| 00 charter.md | roster.md hq.html team/community-supply-lead.md skills/community-supply-lead/SKILL.md <your-project -company/ ├── 00 charter.md thesis, CEO verdict, riskiest assumption, org at a glance ├── 01 findings-report.html research report 5 or 9 panels depending on scope ├── 02 hiring-plan.md org chart + sequenced hire order with reasons ├── 04 pitch-deck.md 10-slide VC pitch deck VC track only ├── hq.html dark-mode command centre — open in any browser ├── roster.md LIVE team registry with health scores — updated every session ├── team/ persona briefs, one per hired agent └── skills/ installable specialist-agent skills, one per role INCUBATOR.md session index at workspace root — how the CEO finds your project /plugin marketplace add afsalali1238/Incubator /plugin install incubator Then: "be the CEO for this — I'm building <your idea " - Download — click the link, then click dist/project-ceo.skill Download raw file the down-arrow icon on GitHub - Open claude.ai https://claude.ai → click your avatar bottom-left → Settings - Go to Skills left sidebar → click Add skill → select the .skill file you downloaded - The skill installs instantly. Start a new conversation and say: "I'm starting a new project, act as CEO" No claude.ai Skills tab?You may be on a plan that doesn't support skills yet. Try Claude Code see above or paste the contents of skills/project-ceo/SKILL.md directly as a system prompt. Or drop the skills/project-ceo/ folder into your Claude Code skills directory if you prefer working from source. New project: - "I'm starting a new project — be the CEO for this." - "I want to build <X , help me kick it off." - "Run point on this build as founder." Returning to a project: - "Let's work on the project." - "CEO check in — where are we?" - "Board meeting." - Just name the project — the CEO finds INCUBATOR.md and picks up. Two modes depending on your setup: Claude Code / Task tool available: agents run as real subagents — the CEO spawns them, they work independently, and return results. Full multi-agent execution. Claude Cowork / standard claude.ai: agents activate as inline personas. The CEO labels them clearly Activating: Growth Lead ... Back to CEO and switches between them in the same conversation. Same domain expertise, no parallel execution. The skill is useful in both modes. The difference is speed and parallelism, not quality of output. Incubator/ ├── .claude-plugin/ │ ├── plugin.json │ └── marketplace.json ├── assets/ │ └── banner.png ├── skills/project-ceo/ │ ├── SKILL.md │ └── references/ │ ├── interview.md decision tree + time-budget question │ ├── research.md method + 20-search hard cap + graveyard protocol │ ├── org-design.md vertical roster patterns + sequencing │ ├── agent-skill-template.md brief + skill templates + quality gate checklist │ ├── starter-pack.md charter, roster, INCUBATOR.md format │ ├── orchestration.md hire/fire/delegate/spawn + board meeting + health scores │ └── hq-template.md the HQ Dashboard HTML template ├── examples/ │ ├── README.md index of all examples │ ├── airbnb/ ★ lead example — all 7 output files + hq.html │ ├── slack-company/ B2B messaging wedge │ ├── uber-company/ ride-hailing city launch │ ├── gymshark-company/ DTC fitness apparel │ ├── oculus-company/ consumer VR hardware │ └── 23andme-company/ consumer DNA testing └── dist/ └── project-ceo.skill packaged installable These are personal Claude Cowork skills. If you have them installed, the CEO uses them automatically. If not, equivalent behaviour is built in as a fallback. — relentless one-question-at-a-time interviewing Phase 1 fallback: built-in decision tree grill-me — surface assumptions, avoid overcomplication built into operating principles karpathy-guidelines — Visual Document format used in Phase 3 fallback: self-contained HTML generated inline vd — autonomous-loop discipline fallback: Search N/20: counter in Phase 2 autoresearch A raw prompt gets you a smart answer. This gets you a persistent company . The difference: Phases enforce order — you can't hire a team before the research exists. The CEO won't let you build before sense-making. Research has a hard cap and accountability — 20 searches, counted out loud, with explicit INSUFFICIENT DATA flags. Raw prompts hallucinate gaps. The roster is derived from evidence — every hire is cited to a specific research panel, not generated from a template. Sessions persist — INCUBATOR.md means the CEO picks up exactly where you left off, without you re-explaining the project. Agent quality is gated — every generated skill goes through a cold-start critic pass before you install it. The skill is useful in a single session. It becomes a different tool over weeks. Long AI sessions accumulate state drift — the model reconstructs your project from context each session, and that reconstruction drifts slightly each time. Three engineering decisions address this: Phase checkpointing. INCUBATOR.md is written the moment the Phase 1 thesis is locked — not at the end. Any session dropout between Phase 1 and Phase 4 is recoverable from disk. session-state.json is written after Phases 2, 3, and 4, carrying phase, confidence, open risks, and a 20-word summary. On resume, the CEO reads this first — not the conversation history. Canonical JSON state. roster.md stores agent team state as a JSON block at the bottom of the file. The CEO reads and writes that JSON block first, then regenerates the markdown table from it. Markdown is presentation only. The two cannot diverge. Cold-start quality gates. After generating each agent skill, the CEO pauses and waits for a CRITIQUE token from the user. This triggers a new generation turn with a cold-start evaluation instruction — the model must evaluate the skill as if it didn't write it. In-stream self-critique asking the model to evaluate in the same output pass it just generated is not real evaluation. The separate turn is. These are prompt-level mitigations, not runtime enforcement. The substrate limits no true parallel agents, no runtime schema enforcement, no live KPI access are documented in the full-thesis writeup /afsalali1238/Incubator/blob/main/docs/full-thesis.html . Built as a founding operating system , not a company simulator. What you get is a structured pre-build research-and-planning ritual with a persistent state layer — the kind of thinking an experienced founding team does before writing a line of code, compressed into a turn-based session. The model plays multiple personas, all from the same context. That's a planning artifact, not a real team. The value isn't the illusion of an org — it's that the phases enforce order: you can't hire before you've researched, can't build before you've validated, can't add Scale roles before you've earned them. Sense-making before building. Agentic coding makes building feel free; that's the trap this is designed to prevent. Roster pattern adapted from slavingia/skills https://github.com/slavingia/skills . MIT © Afsal Ali https://github.com/afsalali1238