ctx-grill

An interview that sharpens a plan until we reach shared understanding. It differs from generic grilling in three ways: existing contexts are read before any question is asked, questions are ranked by whether the answer would change the plan's direction, and resolved decisions are written back so they are never asked again. The interview engine is adapted from grilling in mattpocock/skills (MIT).

Language: mirror the user's language.

Steps

1. Ground: read the context entries relevant to this plan. Sort what they settle into: already decided (never ask — cite the file), assumed but unverified (confirm cheaply), genuinely open (the question pool).
2. Interview: walk down each branch of the decision tree, resolving dependencies between decisions one by one. Ask one question at a time and wait for the answer before the next. This skill stops at diminishing returns rather than emptying the tree, so the question count stays small and batching would buy little against a real cost: one round can pair two questions that turn out to depend on each other. For each question, give your recommended answer, what it rests on, and what would flip it. Do not attach a confidence score; it is not checkable, and it reads as more authoritative than it is. If a fact can be found by exploring the environment (filesystem, tools, docs), look it up rather than asking. The decisions are the user's — put each one to them and wait.
3. Rank as you go: take branches in order of "would the answer change the plan's direction"; within a branch, follow dependency order. Re-rank after each answer — an answer can reshape the tree.
4. Stop at diminishing returns: when no remaining question would change the plan's direction, stop. Name what was left unasked and why it is safe to defer — do not grind every branch to the bottom.
5. Write back: file one-way or durable decisions with /to-ctx the moment they resolve; batch the small ones at the end. Close with the sharpened plan, the decisions filed, and the open items deferred.

Rules

• Do not act on the plan until the user confirms we have reached a shared understanding.
• Never ask what contexts already answer; cite the context file instead.
• Interactive only: if the session cannot receive answers (CI, loops, background runs), refuse and say why.
• Dependency direction: may use /to-ctx; never invokes user-facing skills.