AGENTS.md — Elicitly demos

Guidance for AI coding agents (and humans) changing this repo. Do not create CLAUDE.md; this repo uses the AGENTS.md standard.

What this repo is

Agent Skills (spec) that demonstrate Elicitly's tools with fixed sample data. They are references to copy and adapt, not integrations. The tools themselves are documented at https://www.elicitly.ai/docs/tools/; check a skill's tool calls against that reference, not against memory.

A skill

skills/<name>/ holds SKILL.md plus assets/:

• Frontmatter: name (matches the folder), description, license: MIT-0, and compatibility naming the edition and tools ("Elicitly Free or Pro (elicit_display) …" or "Requires Elicitly Pro (elicit_approval, …)") plus any origin to allow.
• Templates (*.liquid) and sample data (*.json) live in assets/, and SKILL.md tells the agent to pass them verbatim. Keep layout out of the model's hands.
• Pro review tools follow one loop: create → give the user reviewUrl → elicit_await until not pending → elicit_result → act on decision. Never act on a chat claim of approval.
• Dialogs (elicit_confirm, elicit_form) handle all three outcomes: yes, an explicit no, and no answer (null / error). No answer never means yes.

Rules

• Demos only. Every description starts "Elicitly demo". A demo with made-up business data says so ("Elicitly demo (sample data, …)") and ends "Use only when the user asks for … the <name> demo", naming its sample entity (Brightwater, Harbor, Larkspur): it must never trigger on a real-world request and answer it with sample data. (Jupiter, with real public facts, may trigger more broadly.) Each SKILL.md keeps the Demo note under its heading, and the step that would act says "In a real skill, … would happen here."
• Made-up business data. Invent companies, people and figures; never use real customers or real people's names. Images committed here must be openly licensed and credited on the card. The salespeople photos are licensed stock that stays hosted on www.elicitly.ai, not committed here.
• Keep the README in sync. A new or renamed skill updates the demos table (Demo, Tools, Edition, Try saying). Free demos first, and the "Try saying" prompt must match the description's trigger.
• Accessibility (rendered output). WCAG 2.2 AA: text ≥ 4.5:1 contrast, images with alt, keyboard-reachable scrolling regions, and motion-reduce: variants for any animation.
• Review-page templates (displayTemplate, rowTemplate, …) use plain HTML and a <style> block; give links target="_blank". Display templates may load DaisyUI and Tailwind from jsDelivr.
• Conventional Commits (feat:, fix:, docs:, …).
• Never commit or push without explicit user approval. Show the diff first.

Checking a change

• Every SKILL.md frontmatter parses as YAML and name matches its folder. Watch for : (colon, space) inside an unquoted value such as description: YAML reads it as a new key and the skill fails to load. Rephrase, or quote the whole value.
• Every assets/*.json parses.
• Every template renders with LiquidJS against its sample data, with jsTruthy: true and outputEscape: "escape", which Elicitly uses.
• Run the skill once in a host: a display in one that renders MCP Apps, a dialog in one that renders form elicitation (see the Support Matrix), and Pro skills against the hosted server.