intent-interview
Extract the real outcome from a vague ask before any work. Use when a request is broad, multi-step, has unstated decisions, or the user says "let's go deep", "get your context ready", "which skills",
SKILL.mdjcode/skills/intent-interview/SKILL.md
GitHubIntent interview
Turn a vague ask into a written contract, then work the contract. The contract is the artifact; the interview is how it gets written.
Outcome first. The interview exists to make one sentence true: I know what "done" looks like and what I am allowed to assume. Everything below serves that sentence.
The ladder
Run these in order. Stop as soon as the contract is writeable — depth is not the goal.
- Read the ground. Inspect the real thing before asking about it. Files, config, docs, installed skills, command output. A question the filesystem answers is a wasted question, and a wasted question is what makes interviews feel like tax.
- Name the outcome. One sentence: the artifact or the changed behaviour, and who sees it. Draft it from what you read and put it in front of the user. Vague in, vague out; a wrong guess corrected in one line is the cheapest possible repair.
- Find the forks. The open decisions, ranked by how much the finished thing changes if the answers differ. Impact = architecture, file layout, user-visible behaviour, or hard to reverse. Two forks deserve a question. Six deserve a ranked list with recommendations.
- Ask the fork that unblocks the most work. One question, in the user's terms, with a recommended pick and its trade-off. Batch only the forks that share an answer.
- Log the rest. Every fork the user did not answer becomes an assumed decision in the ledger, marked as assumed and reversible. The interview ends; it does not become a questionnaire.
- Write the contract.
OUTCOME / DONE WHEN / ASSUMED / OUT OF SCOPE. Keep it where the work will find it. Then execute.
Question craft
- Ask about the decision, not the topic. "Where does the code live" is a fork. "Tell me about the architecture" is a lecture request.
- Carry a recommendation. Every question ships with a pick. A user who answers "yes" to your proposal is faster than one who authors a proposal.
- Prefer the concrete artifact. Show a shape, a path, a command. "Option A: driver in-tree, one file" beats "Option A: simpler".
- One question per turn for high-impact forks. They earn thinking time.
- Answerable in a sentence beats answerable in an essay. If the answer needs five sentences, the fork was not a fork.
- Absence is data. Silence on a high-impact fork means either it did not matter or the user is tired. Log it, pick, continue. Do not re-ask.
Ask-first, grilled
ask-first lists the forks. This skill does more: it reads the ground first, names
the outcome, asks only the unblocking ones, and writes the contract. When the ask is
already sharp, the ladder collapses to step 6 and the contract is one line.
Output discipline
The contract is short because the interview paid for it. In chat:
- Lead with the answer or the decision, not the reasoning that produced it.
- One table per reply, and only when rows genuinely differ.
- Cap a chat reply at roughly fifteen lines. Deeper material goes behind a pointer, not into the message.
- Reference and prose are written in normal English. Compression registers apply to chat and status lines, where the reader is present to be terse with.
See REFERENCE.md for the question bank, the failure modes, and how
this composes with the decision ledger.