Incubator — a Claude skill that acts as founding CEO

A Claude Code plugin that interviews you into a testable hypothesis, researches the market under a hard search cap, then generates a sequenced team of specialist agent skills.

@afz

github.com/afsalali1238/Incubatorvisit ↗

Incubator — a Claude skill that acts as founding CEO

process.log5 fields

toolsClaude CodeClaude Coworkclaude.ai

iterationsMany attempts

# how_it_was_made

Incubator is a skill you invoke at the start of a project, and it becomes a domain-expert CEO for that specific vertical. It runs seven phases in a fixed order: interview until the idea is falsifiable, research autonomously under a cap of twenty searches counted out loud, publish a findings report that is allowed to tell you to stop, then derive a four-to-seven agent roster where every hire cites the research panel that justified it. I decided the phase ordering and the constraint that you cannot hire before research exists; the AI wrote the reference files, the roster patterns per vertical, and the agent skill templates. The reliability work was the real engineering: session state is checkpointed to disk the moment the hypothesis locks rather than at the end, so a dropout mid-run is recoverable, and the team roster is stored as canonical JSON with the markdown table regenerated from it so presentation and state cannot diverge.

// best_dead_end

questions0

no questions yet

ask how it was made — the dead-ends are usually the good part

ask about the process…sign in

~/slopfolio··

I first had each generated agent skill critique itself in the same output pass that produced it. That is not evaluation — the model just endorses what it wrote. Real quality gating only worked once the critique was forced into a separate generation turn with a cold-start instruction to judge the skill as if it hadn't written it.

# prompting_approach

Phase gates over open-ended prompting — each stage refuses to start until the previous one produced its artifact, which stops the model racing ahead to the fun part. Research got a hard numeric cap with the count surfaced in the output and explicit INSUFFICIENT DATA flags, because an uncapped research loop fills gaps by inventing them. Anything the model generated for install went through a separate-turn critic pass before it counted as done.