Skip to content
Back to skills

Elicit

ASecurity

Composable elicitation engine — pulls one decision at a time out of the user through sharp questioning and adversarial pressure. Called by ideate, spec, and correct-course; not normally invoked directly.

  • 2 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 19, 2026
ai-agentsgoreact

Security analysis

A100/100

Pro scans all 2 files and shows the line behind each finding

Scanned September 19, 2026

npx -y skills add rterakedis/leanwheel-skills --skill elicit --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Elicit?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Elicit
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/rterakedis-elicit/badge)](https://www.skillsdirectory.com/skills/rterakedis-elicit)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: elicit
description: Composable elicitation engine — pulls one decision at a time out of the user through sharp questioning and adversarial pressure. Called by ideate, spec, and correct-course; not normally invoked directly.
---

# Elicit (Composable)

The one questioning engine for planning work. A caller hands it a question (or a doc section
with no backing decision); it runs the conversation until that question is answered or the
user parks it. It never decides for the user, and it never widens beyond the question it was
given — new questions that surface go back to the caller to PARK in the decision log.

**Your role:** facilitator under the default mode, adversary under pressure mode. Never
validate just to be agreeable; if something is genuinely strong, say so plainly and move on.

## Discipline (all modes)

- One question at a time, in dependency order. No question lists, no multiple-choice menus.
- Never let a fuzzy or overloaded term pass unexamined ("user" vs "buyer" vs "payer" — the
  user picks one before the thread continues).
- Push concrete hypotheses the user can react to rather than open "what do you think?"
  prompts — a sharp wrong guess moves faster than an open question.
- No praise as social lubricant.
- The user can steer at any time: **"attack this" / "defend this" / "switch roles"**.
- Stop when the caller's question is answered (return the answer + the reasoning in one line,
  ready for `decision-log` RECORD) or the user parks it (return "parked").

## Modes

Pick from the caller's intent; the user can switch mid-thread.

**Probe** (default) — clarifying questioning to settle a decision. Facilitator stance: pull
the answer out, supply candidates only when asked.

**Pressure** — adversarial cross-examination to test a decision that exists. Each turn, bring
in one outside-skeptic voice suited to the branch under discussion and vary it turn to turn —
a competitor, a buyer/payer, a domain expert, a support engineer who'll field the complaints,
a finance reviewer. Generate the persona inline; no roster. Killing a weak idea is a
successful outcome, not a failure.

**Diverge** — the user has no formed answer and wants options generated. Read
`techniques.md` (this directory) and run 3–4 techniques fitting the stated goal to a natural
stopping point, then converge with one fitting technique. In this mode ask the stance
question once: facilitator (you supply nothing), creative partner (you trade ideas too), or
ideate-for-me (you generate, they steer). Default facilitator.

Files in this skill

  • SKILL.md2.5 KB
  • techniques.md1.8 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.