agentleFS
Alexander-TyagunovAlexander-Tyagunov/magician/skills/conjure/SKILL.md

conjure

Skill14 starsChanged 49 days agoMITView on GitHub

Design dialogue with a visual companion, ending in an approved spec. Use at the start of a feature, before writing code.

3 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 2 are distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do NOT write any code, scaffold any project, or take any implementation action until the user has approved the spec.

    line 15
  2. Distinctive:

    Before the Completion Signal, grade the spec against references/rubric.md.

    line 26
  3. Very specific:

    Read the code, git log, prior specs and research silently, plus .workspace/shared/north-star.md when it exists.

    line 45
  4. Very specific:

    Write <design-dir>/design-notes.md, stop the companion (${CLAUDE_SKILL_DIR}/scripts/vc-stop.sh <design-dir>), say how to reopen the mockup.

    line 65
  5. Very specific:

    Write .workspace/shared/specs/YYYY-MM-DD-<feature>.md (format: spec-format.md), ask *"Spec ready — lock it in?"* (Approve / Request …

    line 69

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/conjure — Design Dialogue

Bundled command: if kg is not found, run it as ${CLAUDE_PLUGIN_ROOT}/tools/kg; if that is missing too, skip the kg steps and search the code directly.

Run a structured design dialogue before writing any code, and produce an approved spec and design artifacts. In every mode, the approaches, architecture and directions are written as Mermaid + HTML under <design-dir> (.workspace/shared/designs/, or .workspace/shared/mockups/ in Design-Only mode), so they outlive the surface that showed them.

<HARD-GATE> Do NOT write any code, scaffold any project, or take any implementation action until the user has approved the spec. This applies regardless of perceived simplicity. </HARD-GATE>

Process — Gated Dialogue

Core rule: each gate is one turn. End your turn after each gate. Do NOT advance until the user replies. Never collapse two gates into one message. Order: Explore → Clarify → GATE 0 → 1 → 2 → 3 (UI only) → 3.5 or 4 → commit. Read a gate's reference when you reach it.

Autonomy: the GATEs (0–4) and the final commit are the only pauses. Around them, exploration runs autonomously: Read/Grep/Glob, kg query and read-only git need no confirmation; a read is never a gate. Re-gate only on real side effects: writing specs/tokens/mockups and git add/commit. Commands this skill does not pre-approve (openssl rand, command -v mmdc, node, the browser opener, WebFetch) prompt as usual. See lore/autonomy.md.

Completion Signal

Before the Completion Signal, grade the spec against references/rubric.md; list unmet items.

"Conjure complete. Spec for <feature> approved and committed to .workspace/shared/specs/ (or, in Design-Only mode, mockups saved to .workspace/shared/mockups/). Screens: <design-dir>/screens/v<n>/; screenshots: <PNG paths per capture.md, or 'none'>. Run /magician:blueprint to create the implementation plan."

References (read on demand)

  • GATE 0 menu, mode → surface map, paths, live loop: references/design-mode.md
  • Explore, Clarify, GATE 1–2 procedures: references/gates.md
  • Architecture as code (.mmd + HTML): references/architecture.md
  • GATE 3 steps 0–5 (UI design, /design-sync): references/ui-design.md
  • Design tokens, variation, light/dark, responsive: references/design-tokens.md
  • Design principles and platform guidelines: references/design-principles.md
  • Brand book template: references/brand-book.md
  • Companion server, screen templates (surface B): references/visual-companion.md
  • Screenshots as evidence (capture, diff v<n> vs v<n+1>): references/capture.md
  • GATE 3.5, GATE 4, commit, sharing: references/after-approval.md; spec file format: references/spec-format.md

Steps 1–2 — Explore, Clarify

Read the code, git log, prior specs and research silently, plus .workspace/shared/north-star.md when it exists; then ask one clarifying question per turn until purpose, constraints, success criteria and edge cases are clear (details in gates.md).

GATE 0 — Design Mode

Call AskUserQuestion with exactly the four options in references/design-mode.md: Visual + Strict, Visual + Reference, Text only, Design Only (Visual). End your turn. Visual options use one of two surfaces: (A) /design artboards — Claude Code's bundled /design skill, a research preview (v2.1.234+, Pro/Max/Team/Enterprise), used when the session has the Artifact tool (detect by tool presence); (B) the local companion — ${CLAUDE_SKILL_DIR}/scripts/vc-start.sh, works offline and in CI, the documented fallback. When both are available ask once "artboards or local companion?"; when one is, use it and say so; when neither is, continue as (C) text. Design Only is a flag on A or B (no spec). Do not propose approaches until the user picks a mode. "Skip" switches to text and stops the companion.

GATE 1 — Approach Selection

Present 2–3 approaches: write <design-dir>/screens/v1/approaches.html in every mode, then show it on the chosen surface or describe it and ask with AskUserQuestion. End your turn. Do not present architecture until the user confirms an approach (references/gates.md). On the choice, record it in .workspace/shared/decisions/ (format in gates.md).

GATE 2 — Architecture Review

For the chosen approach only: write architecture.mmd (C4-style component diagram + a sequence diagram of the main flow), render architecture.svg only when mmdc is already on PATH (never install it), embed it in architecture.html (else a pre-rendered box diagram, no CDN), then show or describe it and ask Approve / Request changes. End your turn; advance only on explicit approval (references/architecture.md).

GATE 3 — UI Design (skip if no UI involved)

A design system, not a one-off mockup: read references/design-tokens.md, then follow references/ui-design.md step by step. Step 0 is mandatory: ground in the current design (brand.md, design-tokens.css, the app's styles and components; it is the constraint unless the user asked for a rework or the app is greenfield) and in the platform guidelines (lore/deep/design-principles.md). Step 1: three directions inside those constraints. Steps 2–4: viewports; emit design-tokens.css + brand.md; ONE mockup (light+dark, responsive) after the review checklist, captured when approved (references/capture.md). Step 5: offer /design-sync once; never auto-run it.

GATE 3.5 — Design-Only Close (DESIGN_ONLY only)

Write <design-dir>/design-notes.md, stop the companion (${CLAUDE_SKILL_DIR}/scripts/vc-stop.sh <design-dir>), say how to reopen the mockup. No spec, no plan (after-approval.md).

GATE 4 — Spec Approval, then commit

Write .workspace/shared/specs/YYYY-MM-DD-<feature>.md (format: spec-format.md), ask "Spec ready — lock it in?" (Approve / Request changes), end your turn. On approval: git add the spec (asks), git commit -m "docs: add spec for <feature>", stop the companion, point to /magician:blueprint (references/after-approval.md).

Design principles: one responsibility per unit, clear interfaces, small files, YAGNI, testable components.

Obstacles

A human-in-the-loop dialogue — every gate needs the user — so it reports to the human directly, not upward to an orchestrator, and emits no upward Obstacles block: say any blocker (the companion, /design or Playwright not starting, unreadable brand assets, a missing input) in prose and stop. See lore/obstacles.md.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.