elicit - SKILL.md Agent Skill

name: elicit description: "Extract tacit knowledge through non-obvious microquestions — things only the user knows that can't be found in code, docs, or research. Use when you're about to guess at intent, context, or constraints during SAFEWORD's understanding flow. Also use when user says 'ask me', 'what do you need to know', or when another skill (bdd, brainstorm, debug) needs user context before proceeding. Do NOT use for questions answerable by reading the codebase or searching the web." allowed-tools: '*'

Elicit: Tacit Knowledge Extraction

Draw out what the user knows and you can't research. Then check the latest evidence. Then explore options together.

Iron Law: NEVER ASK A QUESTION YOU COULD ANSWER YOURSELF. Read the code, search the docs, check git history first. Only ask what's left.

When to Run

Explicitly — user invokes /elicit
From other skills — any skill can call "run /elicit" before a decision point (e.g. bdd before scenarios, brainstorm when narrowing, debug when symptoms are ambiguous)
Proactively — when you're about to make an assumption only the user can confirm

Relationship to other skills

brainstorm widens the option space (divergent). elicit narrows your uncertainty about user intent (convergent). Use brainstorm when you don't know the options; use elicit when you don't know the user's priorities among options.
SAFEWORD understanding flow already does contribute-before-asking. elicit is the structured form for decision points where guessing would be costly.

Phase 1: Microquestions

Ask one question at a time. Multiple choice (3-5 options). Force yourself to pre-think the option space so the user reacts rather than generates.

Ask the most discriminating question first — the one whose answer most narrows what you'd do next (highest information gain). Don't walk the categories in order; pick the probe that eliminates the most uncertainty. LLMs default to low-information questions — counter that deliberately. (This is the convergent mirror of debug's disconfirm-first.)

What to ask about

Intent: Why this way and not another? What was tried and rejected?
Constraints: What's off-limits? What's coming soon that would change the answer?
Priorities: When two good options conflict, which value wins?
Audience: Who reads/uses this? What do they already know?
Omissions: Is something missing because it's not ready, not wanted, or not thought of yet?

Question format

Good:

This could live in (a) the onboarding flow, (b) the decision-making page, or (c) as a standalone guide. Where would someone look for it?

Bad:

Where do you think this should go?

The good version does the thinking. The bad version makes the user do it.

Guard against anchoring. Pre-supplied options cut effort but also lead — the user may pick from your frame and skip the answer you didn't list. Keep options non-leading and roughly mutually exclusive, and always offer a real escape ("none of these / something else") so the frame itself is rejectable.

How many

Keep asking until one of:

Answers start repeating or converging
You have enough to distinguish between options in Phase 3
User says "enough" or "go"

Typical: 3-7 questions. Fewer for small scope, more for ambiguous scope.

Phase 2: Research

After elicitation, check the latest evidence on the decision at hand. Web-search for:

Meta-analyses or systematic reviews
Replicated findings with effect sizes
How leading organizations handle this
Boundary conditions and known limitations

Skip if the task is purely operational (no empirical claims involved). Defer if the calling skill has its own research phase — don't duplicate work.

Phase 3: Options

Present 2-3 options. Each option:

What: the approach, concretely
Why it's good: what it gets right
What it costs: complexity, length, tradeoff
Evidence: what supports it (from Phase 2, if applicable)

For small scope (patch/task): just do it — skip the proposal, apply the best option directly. For larger scope (feature/policy): present options, let the user pick.

From Other Skills

When called from another skill, the elicitation results feed directly into the calling skill's next phase. No separate artifact — the conversation is the artifact.