canvas-design

Fires when

• ▸user wants to generate a poster, plate, zine cover, or other static visual artwork
• ▸user wants Claude to author a design philosophy first and then express it on a canvas
• ▸user wants a single-page PDF or PNG with a strong typographic gesture and sparse marginalia
• ▸user wants design-taste guidance (philosophy-first, minimal text, monumental gesture, subtle reference)
• ▸user wants access to a curated typeface library without sourcing fonts themselves

Skip when

• ✕user wants pixel-precise layout from a brief — this skill resists templates, it teaches an aesthetic
• ✕user wants raster illustration, photo composites, or any output that depends on a pre-trained image model
• ✕user wants logo design with brand constraints — the skill's "subtle reference" gesture pulls against literal brand marks
• ✕user wants a one-shot fill-in-the-blanks generator — there is no make-my-canvas.py, the model must pick a renderer and write the code
• ✕user needs deterministic output across runs — different models or different days will produce different compositions

Takes

• prompt any short brief; the skill expects a "subtle conceptual thread" (a niche reference) rather than a literal subject
• text:markdown optional — the skill itself authors a philosophy.md as step 1; users do not supply it

Returns

• file:pdf vector geometry + bundled TTFs render in ~60ms at A2; quality is bounded by the model's design taste, not the skill's tooling
• file:png same composition rasterised; requires a separate renderer (sips/pdftoppm/Pillow) — not part of the skill itself
• file:markdown the design philosophy itself, ~5-6 paragraphs; useful as a reusable aesthetic brief independent of the visual output

Install

pip install reportlab Pillow

• macos: sips (built-in) or brew install poppler — for rasterising the PDF to PNG
• notes: SKILL.md declares no renderer; the model picks one. reportlab is the natural fit for vector + TTF registration. The 54 bundled TTFs in canvas-fonts/ are the only first-party tooling.

Zero scaffolding. No requirements.txt, no package.json, no example script. Required deps must be inferred from the visual the model wants to make. Pillow alone is not enough — it cannot embed TTF subsets into a PDF.

Caveats

• Zero scaffolding ships with the skill. No requirements.txt, no example script, no make-my-canvas.py — SKILL.md never names a renderer (reportlab, cairo, Pillow, matplotlib are all unstated). The model must pick one and write the drawing code from scratch every run.
• Font registration is silently fragile. reportlab raises KeyError if you call setFont("BigShoulders", 80) without first registering the TTF — no graceful fallback to Helvetica. Any agent that copies font names from the SKILL.md cookbook without also discovering canvas-fonts/ will crash, not degrade.
• Output quality is bounded by the model's design taste, not the skill's tooling. The skill teaches *what* to make (philosophy-first, minimal text, monumental gesture, subtle reference); it does not teach *how*. A weaker model will produce a weaker poster even with the same philosophy.
• The "FINAL STEP" injection ("It isn't perfect enough. It must be pristine, a masterpiece") is a hidden second-pass prompt that pushes the model toward more restraint, not more invention. Watch for compositions that ratchet themselves toward emptiness — the memorial plate in this review is what that ceiling looks like at its honest extreme.
• PDF to PNG is not part of the skill. SKILL.md says "Output the final result as a single, downloadable .pdf or .png file" but provides no path from one to the other. macOS sips works for single-page PDFs only; multi-page exports need pdftoppm or Pillow.
• The "subtle reference" instruction is load-bearing and undocumented in its risk: if the model picks too literal a reference (e.g. a real artist's signature motif), the skill silently violates its own "never copying existing artists' work" warning. We chose a thematic reference (Messner's philosophy of honorable retreat — "the mountain is preserved by those who turn back") that is never named in the work itself; not all model runs will be this careful.
• Deterministic geometry is the user's responsibility. The skill's aesthetic asks for "perfect shapes" and "repeating patterns" but does not warn that procedural noise (random.uniform, time.time seeds) will produce a different poster every run. Seed your randomness or your "masterpiece" is unreproducible.