trainingsites/campus-ai-os/grill-me

grill-me — interview me until there are no gaps

The extraction engine for the business-memory moat. Where the council converges one idea to a choice, grill-me does the opposite: exhaustively pulls what's in your head into structured, reusable context. Adapted from Matt Pocock's grill-me; capture pattern per Nate Herk.

Kernel preamble §4 applies: references/kernel-preamble.md.

Kernel preamble §5 applies: references/kernel-preamble.md.

CRITICAL DIRECTIVES

• One question at a time. Ask → get the answer → ask the next. Never batch, never form-dump.
• For each question, offer your recommended answer the user can accept or override (lowers effort, speeds the grill).
• Context-first — collect as much as possible, ask as little as needed. Before asking, check what's already known (resolved shared/* via CONTEXT_CANON (.campus-os/context-files.md, else the plugin's templates/context-files.md), the wiki/ index, dean/memory.md, connected tools). If the answer is already on file, state it and confirm instead of asking cold.
• Relentless until shared understanding. Keep going (5 questions or 50) walking each branch of the decision tree, resolving dependencies in order, until there are no holes. Propose when you think it's complete; the user decides when to stop.
• Never fabricate. Flag instead. When the user doesn't know something (it lives with another person/tool/file), record an open flag ("get X from Y, then re-grill") rather than guessing.
• Tool-agnostic; zero hardcoded paths.

Step 1 — Frame the target

Name what we're extracting (a process, an offer, a launch, "the whole business"). Inherit the run_id if one exists (council/brief); else have the orchestrator mint one. Immediately call session-capture to open the Tier-1 log so nothing is lost from question one.

Step 2 — Explore before asking

Pull everything already captured on this target from resolved context, wiki, memory, and connected tools. Summarize what you already know back to the user and confirm it. Only the gaps become questions.

Step 3 — Grill the decision tree

Walk branch by branch. One question, your recommended answer, their response. After every answer, checkpoint to the Tier-1 log via `session-capture` (discovery notes · key decisions · Q&A log · open flags). Follow dependencies — don't jump randomly. Surface open flags as they arise.

Step 4 — Confirm completion

When branches bottom out and you and the user share the same understanding, say so and ask to wrap. If gaps remain that only an external source can fill, list the open flags and what to fetch.

Step 5 — Persist + route (via session-capture)

Trigger the Tier-2 twin: the distilled wiki entry + the `context-record.json` (gate fields default private/internal) + manifest upsert. Then offer to route confirmed facts to the file each concept already resolves to (tools/resolve_context.py --concept icp; on a founded campus that is audience.md, never a new icp.md beside it; offers to the resolved offers file; the manager's own entry file — resolve ENTRY_FILE per PLATFORM.md, never assume a filename) and to update any skill/doc the grill exposed as thin (e.g., "your packaging skill is missing this nuance — update it?"). Offer and wait — never auto-apply.

Re-grill

"Grill me again on {slug} — here's what changed" loads the prior log + record and extends it (version bumped), never starts cold.