Intake Interview Discipline

Intake Interview Discipline

Internal shared skill. Single source of truth for interactive-grade question depth in unattended long-run hosts: everything only the human can answer is asked up front while they are present; everything the run answers alone is recorded for their return. Composes clarifying-question-discipline (how to ask) and decision-fog-discipline (when understanding is clear enough); adds what unattended hosts lack — approach freeze, runtime self-answer rules, and the acceptance ledger.

Prerequisite check: this skill declares frontmatter dependencies; verify each is available on load — if one is missing, abort and print the install command (npx skills add FuDesign2008/open-skills -g --skill <name> --yes). No silent fallback.

Iron rules

1. Fog-bounded depth, not round-bounded — interview until no blocking fog remains (graduation per decision-fog-discipline), never a fixed questionnaire. A clear task graduates fast; a foggy one earns more rounds. The human's time is spent on decisions only — facts get investigated, not asked.
2. The frozen approach is the run's compass — once frozen into the contract, the approach MUST NOT be silently replaced mid-run. If run evidence falsifies it: clean stop + ticket report (§B). A visible stop is success discipline; a silent pivot is failure.
3. Every runtime self-answer is an assumption until reviewed — recorded in the ledger with rationale and impact-if-wrong; the human MAY overturn any entry at acceptance.

A. Intake deep interview (human present)

Presence tiers — interview depth follows human availability, because a question is only cheap while someone can answer it:

• Present (default) — resolve open decisions as per-decision questions until fog graduates. Self-answers are reserved for explicitly low-impact items; anything whose impact-if-wrong is high (positioning that trusts an upstream MR, verification strategy, external contracts) gets asked, not defaulted. Open the intake output with a three-part base — restatement / key elements / open questions — so the human checks the AI's unfolded understanding before anything is frozen into a compact contract.
• Declared or structural absence (「无人值守」/ "I'm off", batch child runs, scheduled pulls, card-inherited intake) — the once-confirm + full-ledger mode applies unchanged: freeze what investigation and conservative defaults support, flag everything as assumptions, close the loop at ledger review when the human returns.
• AI proxy (opt-in third source) — when the host run's Stage-exit policy is ai-proxy (per ai-proxy-discipline), present-mode methodology runs with the proxy as answer source at the enumerated checkpoints, every answer ledger-marked proxy-made. Reserved-list items stay human-only at every touchpoint.

1. Destination — one short success picture (what "done" looks like).
2. Fog map & tickets — list open decisions as tickets per decision-fog-discipline; keep the list short; mark blocking vs deferrable.
3. Resolve one ticket per turn — clarifying-question-discipline (one question per turn, recommended answer required); facts are investigated, not asked.
4. Approach comparison — 2–5 approaches, one comparison table (core idea / files touched / pros / cons / complexity / recommendation); the human picks. Lean path with no fog: 1 approach + a risk note is enough.
5. Freeze — write into the host contract: chosen approach (one line + pointer to the table), resolved tickets, deferred tickets (with reason), initial assumptions, plus a Decisions-I-made-for-you section — every self-answered decision with its impact tier, so the approval event shows what was decided without the human. Mark each entry factual (branch baselines, dependency existence, target paths) or preference; unmarked defaults to factual, and hosts verify factual entries at their dispatch boundary. Consequence descriptions MUST be derived by walking the enqueue configuration → child default behavior path as actually configured — e.g. proxy off + child manual mode means stage exits will stop and ask the user, not "follow frozen decisions"; a mis-described consequence is worse than none. Before closing, re-check each self-answered item's impact-if-wrong and escalate any high-impact item to a question now: this is the last cheap moment to ask. Templates in reference.md.
6. Bounded pre-launch self-review — one checklist-level pass over the frozen contract: covers the destination? risks named? budget + constraints present? anything self-contradictory? Fix gaps and re-check once; a remaining blocking doubt goes back to the human now — they are still present; after launch there is no one to ask.

B. Runtime self-answer (human absent)

When a decision arises mid-run, take the first branch that closes it:

Batch/queue hosts: a clean stop on one task never cancels the others (non-blocking failure).

C. Acceptance ledger

• One ledger per run, carried in the host's contract artifact (task card / run contract / completion report); fields in reference.md.
• The host's completion report MUST surface for human review: unresolved tickets, low-confidence assumptions, and every ledger entry whose impact-if-wrong is high.
• Acceptance findings feed the next run's intake — interview quality compounds across runs.

Integration guide (for referencing hosts)

1. Declare intake-interview-discipline in frontmatter dependencies; abort at startup if missing.
2. Keep exactly these touchpoints in the host body — do not restate this skill's prose:
   • Intake stage: one pointer line (deep interview with presence tiers + freeze + impact re-review + bounded pre-launch self-review per this skill)
   • Contract fields: point at reference.md templates (frozen approach, tickets, assumptions, Decisions-I-made-for-you section)
   • Run stage: one line — self-answer per this skill (impact-gated ladder); falsified approach → clean stop, no silent pivot
   • Report/acceptance stage: one line — ledger surfaces unresolved + high-impact entries for human judgment
3. Attended hosts that pause at every gate by design (solve-workflow-style) do not need this skill — their humans are present. This skill exists for hosts whose value promise is that nobody has to be present mid-run.
4. Auto-mode intake with no human present at all: run §A in self-answer mode — investigate, propose defaults, freeze them flagged as assumptions; the "interview" then happens when the human returns, via ledger review.

Forbidden

• Restating clarifying-question-discipline or decision-fog-discipline rules in host bodies (thin pointers only)
• Fixed question counts or questionnaire dumps ("ask exactly N questions")
• Silent mid-run approach changes; treating budget exhaustion as success
• Marking outcome-type standards as passed by ledger self-assessment (humans judge outcomes)