Glyph

Glyph

Render the answer as ASCII art plus semantic emojis, inline, immediately. All output renders in a monospace terminal with no external tools.

Core principle: Encode information into structure, not decoration. Every diagram element should communicate something meaningful.

Execution (run this, do not ask first)

A human asking "how do I use this" gets a URL, not a paraphrase: https://orchestkit.yonyon.ai/docs/reference/skills/glyph#examples (what it draws, the two picks, three real renders, the exact invocation; generated from this skill's examples/_featured.md, so it cannot drift from the skill).

The whole point is speed, so there is no setup phase.

Already loaded means render, not hand off. Once this file is loaded, draw the answer here. Do not call the Skill tool for glyph, /glyph, or /ork:glyph. A second invocation loads a skill again and re-runs the planning step. The 2026-09-15 Devin transcript showed "Invoked skill glyph" then "Invoked skill ork:glyph", one render stacked on another. Whether that re-entry is also why the Thoughts text repeated is not verified here; do not re-enter either way. The front door does not delegate. A user may have a personal skill also named glyph (/glyph); it lives in a separate namespace, may call this skill for its chat route, and is never called from here (references/dials.md, "Personal skills named glyph").

With no argument, the topic is the current conversation. Measured over a real 13-prompt session: zero asks supplied a self-contained topic, and the one direct invocation passed nothing at all. glyph on its own means "render where we are right now": the open work, the decision just reached, the state of the thing being discussed. Render that; do not ask what to draw.

Given a topic (or the conversation, when none is given):

0. Resolve the two dials, then announce them in one line. Surface is chat, ask or page (default chat); audience is operator or novice (default operator). A flag beats inference: --chat, --ask, --page, --eli5, --decide, --signoff. Print exactly → chat · operator (surface, then audience) and a blank line before the answer. Inference rules, the ask guardrail and the --decide shapes: Read("references/dials.md").

1. Render immediately. Do NOT call AskUserQuestion to pick a format (the ask surface and --signoff are the only questions this skill asks), do NOT call TaskCreate, do NOT spawn an Agent. Choose the form yourself from the topic shape and draw it. Asking first defeats the skill.

2. Pick the form from the shape of the data, using the pattern library below:

   Topic shape	Form
   inventory, audit, budget, "what is using X"	inventory render (templates/inventory.md): header meter, traffic-light sections, icon / label / bar / value / action rows, totals, arithmetic summary, caveats
   state / progress / health	status box + bar meters
   A vs B, options, trade-offs	comparison table or side-by-side boxes (one narrow table in a non-TTY surface)
   steps, pipeline, hand-offs	left-to-right flow with ──▶
   containment, layers, layout	nested boxes / tree
   ranked list, scores, counts	table + bar meters
   over time	sparkline or milestone track
   triage open issues, what is left, categorize the backlog	triage page (templates/triage.html). Not an inline render.
   explainer page, NOW vs IDEAL, "show it moving" (page route only)	JS art piece on the page (references/js-art.md). Chat stays ASCII.

3. Emit inline in the reply, except the triage page. Never write a file unless the user asked for one. The triage row above is the exception: that answer is templates/triage.html, not an inline render. Every other answer stays in the reply.

4. Use the closed status set, the domain-icon legend and the box-drawing vocabulary defined in rules/visual-style.md (shipped with this skill) and tokens.json (icons.*). Status icons pair with a word (✅ done, ❌ failed, ⚠️ warning, 🔴 high risk, ℹ️ caveat). Domain icons (🐳 docker, 📦 package, 🧪 test, ...) go one per row in the leading column so they scan as a legend; never inside a prose sentence, never in chains.

5. Stay inside the budget: ONE render per reply, up to about 50 lines, every line 76 cells or fewer. Two competing renders in one reply is flooding. Above 50 lines it is a page, not a chat answer. The 2026-08-09 budget (12 lines, 40 percent) over-corrected and was reset by operator word on 2026-09-17; the history is in rules/visual-style.md.

6. Verdict first, then the render, then what the numbers told me. One prose line states the point before the render. After the render, two or three bold-led bullets say what is not obvious from the chart. If the reader has to parse the render to find out what happened, the reply failed.

7. Stay honest. If a number is unknown, print ? rather than inventing one. A confident-looking chart built on guesses is worse than prose.

8. Match the width to the host. A render that fits a terminal wraps into a wall in a host that reflows text: CI logs, chat widgets, VS Code chat, web transcripts, agent desktops. Read the surface from the environment the same way you already infer audience and surface: with a terminal (TTY), render as below; without one, cap every diagram line at 72 columns, and render key/value or comparison data as one narrow table or a vertical list, never side-by-side columns. The line budget does not change. Measured failure the rule prevents: a side-by-side key/value board for 5 rows lands at 113 columns, and scripts/render-ascii.sh key-value renders the narrow form of the same data (GH-4159).

When NOT to use this skill: if the deliverable is a multi-section HTML playground, a persisted plan artifact, or any file output other than the triage page, use visualize-plan instead. Glyph is the cheap inline path; visualize-plan is the full pipeline. The triage page stays here.

Triage is a page, every time. Asks shaped like "triage open issues", "what is left", or "categorize this for me" use templates/triage.html and no other layout. This overrides the inline-only rule and the visualize-plan handoff: the page is glyph's, then /page-serve PATH. Fill issues from one snapshot, gh issue list --state open --json number,title,labels,milestone. Put every gh-sourced string in the hidden snapshot textarea as one base64 blob of the JSON, not as raw JSON and not in the lane markup or the raw dump. A title that contains </textarea must never appear as raw markup inside that textarea, or it breaks out before escapeHtml runs. The page decodes the blob, then calls escapeHtml when it inserts each lane item and the raw dump, so a title stays literal text. The open-PRs tile is not in that snapshot. Fill it from gh pr list --state open --json number, and if that command was not run print ?, never a guessed count. Keep the six parts in order: KPI strip, lanes, a route tag on every issue (devin, ork:NAME, hq-ext, 21st-dev, human, external), capability map, DECIDE block, collapsed raw snapshot. Hand the file over with /page-serve PATH. Never paste a bare path. The copy button must emit wave=LANE followups=CSV. When no follow-up is checked, that is followups= with nothing after the equals, never followups=0.

Over budget is the same signal. If the honest rendering needs more than ~50 lines, that is not a bigger chat answer, it is a different deliverable: write the playground or file, then hand the human a URL with /page-serve PATH (a port-free https://<name>.localhost/ route, with a stop) instead of a bare file path or a hand-started python3 -m http.server. Keep a 10-line excerpt in chat next to the URL. The old escape hatch fired on artifact TYPE only, so an over-budget inline reply never tripped it.

JS art is for pages; chat stays ASCII. When the answer is already a page (an explainer, a NOW vs IDEAL plan), the page carries one animated piece above its first beat: an inline <canvas>, a Now/Ideal toggle, a Play/Pause button, and a text alt on the canvas. Build it from references/js-art.md: paste one pattern (flow dots, gate, before/after meter, or real-data bars from a JSON array in the page), then the kernel. draw(ctx, t, mode) is pure, so window.seek(t) can capture any frame to mp4 or gif; prefers-reduced-motion gets one still frame and no animation loop; no Math.random, no network. A piece that shows measured numbers draws ? for a value you do not have. Never put JS art in a chat reply.