pixel-face
Skill393 starsChanged 15 days agoApache-2.0View on GitHub
Create or improve a 32×32 pixel-art persona face for the Frontier Faces demo (examples/04-frontier-faces/personas.mjs) — the animated agent avatars rendered by face-term.mjs / the browser <cotal-face> / the gallery. Use when the user asks to add a new persona/face/avatar, improve an existing one (hair, eyes, likeness, expression), make it "look more like" someone, or fix a face that "doesn't read". Drives the example's own tools (img2rows.mjs, render-png.mjs) and follows FACE-DESIGN.md; uses a multi-agent variant loop for likeness.
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- cold-reviewSkill
- parallel-feature-managersSkill
- cotal-meshSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
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
- Very specific: line 63
Read its entry in personas.mjs + its reference assets/<key>.<ext> and the current render.
- Distinctive: line 70
When matching a real person, generate options instead of guessing:
- Very specific: line 81
Always review visually — the render-png contact sheet and the gallery.html original-vs-pixel view — never blind.
- Very specific: line 84
Don't commit assets/ (reference photos stay local).
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
pixel-face — author & refine the Frontier Faces pixel art
A persona face is pure data in examples/04-frontier-faces/personas.mjs (one entry per
persona). The same data drives the terminal renderer (face-term.mjs), the browser element
(web/cotal-face.js), and the PNG tool — so edit personas.mjs once and it's live everywhere.
Always work from examples/04-frontier-faces/.
Where things are
personas.mjs— the entries (rows, colors, glow, mouths, expr, eyes, lines). Single source of truth.FACE-DESIGN.md— the full conventions (grid zones, color keys, eye recipe, visemes). Read it first.tools/img2rows.mjs— rough a reference photo into a 32×32rows+palette.tools/render-png.mjs— render a 7-state contact sheet (PNG) for review. This is the validator.tools/face-template.mjs— copy-me starter entry that encodes the conventions.gallery.html(all personas, original-vs-pixel) andpreview.html?p=<key>(one persona) — served viapython3 -m http.serverortools/serve-wall.mjs; for visual QA.assets/<persona-key>.<ext>— reference photos. Gitignored / local-only — never commit them.
Persona key ≠ agent name. The roster uses agent-file basenames; the face comes from each agent's
face: frontmatter: elon→musk, steve→jobs, rayan→ray (others match). Persona keys:
neon david ray sven garry jobs musk dario mira bernie michelle.
Conventions (condensed — full detail in FACE-DESIGN.md)
- Grid 32×32,
.= transparent, face centered ~col 15. Zones: hair 1–9 · brows 10–11 (drawn byexpr, notrows) · eyes 12–13 (drawn byeyes()) · nose 14–16 · mouth 18–20 (drawn bymouths) · chin 21–24 · collar 26–31. - Color keys: skin
Plight /Smid /sshadow ·Epupil ·W/wcatchlight ·Rlip ·mlip-shadow ·bbrow ·hhair ·Lcyan +Mmagenta = the cotal glow ribbon (glow: { L: 8, M: 6 }blurs them). Userng(row, c1, c2, key)(exported from personas.mjs) for runs. - Eye recipe: 2-wide iris (
E) on the iris row; row below = catchlight (W/w, OUTER corner)- lid-shadow (
s, INNER).jobsis the "alive" benchmark — match its eye, don't use flat black blocks.
- lid-shadow (
- Surprise = raised brows + open mouth, eyes stay
open(growing the iris behind glasses looks wrong). Don't reintroduce awideeye style. - The face renders at a fixed 32×16 (not scaled to the pane), so faces look the same size as long as the pane is ≥ ~32 cols.
Tools (exact)
# rough a reference photo → 32×32 rows + palette (macOS: uses sips)
node tools/img2rows.mjs --in assets/<key>.png --w 24 --h 30 [--colors 12] [--x0 4] [--y0 0]
# render a 7-state contact sheet (neutral·happy·sad·angry·surprised·talking·blink) to review
node tools/render-png.mjs --file <draft>.mjs --out /tmp/<key>.png --scale 8 # draft: export const entry = {…}
node tools/render-png.mjs --file personas.mjs --persona <key> --out /tmp/<key>.png --scale 8 # existingA draft module must export const entry = {…} and (if it uses rng in expr/mouths)
import { rng } from '/abs/path/examples/04-frontier-faces/personas.mjs'.
After every render, Read the PNG and compare it to assets/<key>.<ext>. Iterate.
Create a new persona
- Pick a
key. If you have a photo, save it asassets/<key>.pngandimg2rowsit for a rough base; otherwise copytools/face-template.mjs. - Hand-edit a
/tmp/<key>-draft.mjs(export const entry) following the conventions above. render-png→ Read the PNG → compare to the reference → iterate until all 7 states read well.- Splice the finished entry into
personas.mjs(keep the file's style; reuse color keys). - Review in the gallery; confirm no other persona changed.
Improve an existing persona
- Read its entry in
personas.mjs+ its referenceassets/<key>.<ext>and the current render. - Edit only the in-scope rows (e.g. hair = rows 1–9). Preserve
eyes(), glasses/beard,expr,mouths,colorsof other features, and the collar/glow — change the minimum. render-pngthe draft → Read → compare → iterate (≥3 passes). Then splice the rows back intopersonas.mjsand confirm the diff is limited to that entry.
Multi-agent variant loop (use for likeness — "doesn't look like them")
When matching a real person, generate options instead of guessing:
- Spawn ~3 subagents on the main working tree (NOT git worktrees —
assets/is gitignored, so reference photos are absent from a fresh worktree). Give each: the reference path, the current entry, the conventions, and a distinct direction. - Each agent writes its OWN
/tmp/<key>-<n>.mjsdraft (never editspersonas.mjs— avoids conflicts), renders it, Reads the PNG, compares to the photo, iterates, and returns the finalrows. - You pick the best, splice it in, review in the gallery. Refine one axis per pass (e.g. head shape, then hair). This is the loop used for garry/musk.
QA bar
- Always review visually — the
render-pngcontact sheet and thegallery.htmloriginal-vs-pixel view — never blind. Judge: does it read / look like the person across all 5 expressions + blink? jobsis the reference standard. Keep faces front-facing/symmetric unless intentional.- Don't commit
assets/(reference photos stay local). Persona edits topersonas.mjsARE committable.
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
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.

