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.