forge-onboard

Forge Onboard — build the Project Brain

A brand-new Forge project has skills wired and a pipeline preset applied, but its "brain" is empty: nothing has read the actual repo and turned it into projectFacts / knowledge entries / seed memory / a tuned pipeline config. This skill is that first conversation. It runs once, at the start of a fresh chat session, and drives a short human-in-the-loop interview to fill that gap — never by guessing silently and writing, always by proposing and waiting for a yes.

Hard rules (non-negotiable)

1. One question per turn. When you need to ask the human something, ask exactly ONE question, then END YOUR TURN immediately — no follow-up questions stacked in the same message, no "while I wait, let me also ask...". The reply arrives as the next turn (the session resumes automatically); you pick up the conversation from there.
2. Never write without an explicit confirm. Before any forge_knowledge upsert, forge_memory write, forge_config projectFacts/pipelineConfig patch, or pipeline-config change, present a short plain-English summary of exactly what you intend to write and ask "Write this?" (a single question, per rule 1). Only write after an unambiguous yes. A "no", a follow-up question, or the user closing the tab must never leave a partial write behind — nothing you propose exists until it is confirmed.
3. Acknowledge every write. Immediately after a confirmed write succeeds, say so in one line ("Saved: "). If a write fails, say so plainly and offer to retry — never fail silently and move on as if it worked.
4. No dead ends. At any point the human can decline a proposal, ask something unrelated, or just stop — the conversation keeps working (or ends cleanly) either way; it never gets stuck waiting on a write that will never be confirmed.

Flow

1. Survey the repo silently (no questions yet)

Before asking anything, spend this first turn reading what the repo already tells you — the goal is to minimize what you have to ask a human. Useful sources, whatever exists: