Are you the author? Sign in to claim
Talisman — turn a codebase into a beautiful, brand-styled PDF guide (LaTeX→PDF). An agent skill that works with any codi
An agent playbook that reads your app and writes a polished LaTeX → PDF guide — with a designed cover, your real colors and logo, native diagrams, and a subtle branded background. Works with any coding AI agent.
npx skills add atexy4ba/talisman-skill
What it does · Install · Usage · How it works · Structure · Requirements
Point Talisman at a repository and it produces a docs/guide/ LaTeX project that
builds to a print-ready main.pdf.
| 🎯 Grounded in the real code | Describes what the app actually does — read from the schema, modules, and flows — not the aspirational claims in a marketing README. |
| 🎨 On-brand | Extracts the product's real palette (even modern oklch() CSS), its logo, and the character of its fonts, so the document looks like it came from the product. |
| 📐 Designed, not dumped | A full-bleed cover page, colored headings in a geometric display font, a slim branded page background, and callout boxes for notes and caveats. |
| 🧩 Illustrated | Native TikZ diagrams — domain models, step-by-step flows, and state machines — with no Graphviz/Mermaid dependency. |
Why "talisman"? A talisman is a crafted object that carries the essence of the thing it represents. This skill distills a whole codebase into a single artifact you can hold.
Talisman is on skills.sh. The
skills CLI detects your agent and
installs it in the right place automatically:
npx skills add atexy4ba/talisman-skill
That's it — works with Claude Code, Cursor, Codex, Windsurf, and every other agent
the CLI supports. Manage it with npx skills list / update / remove, and find
it any time with npx skills find talisman.
Talisman is also just a portable playbook — plain Markdown instructions
(AGENTS.md / SKILL.md) plus Python scripts and LaTeX
templates, with no runtime and no agent lock-in. If you'd rather wire it in by
hand, clone it:
git clone https://github.com/atexy4ba/talisman-skill.git
| Agent | How to enable |
|---|---|
Any AGENTS.md-aware agent (OpenAI Codex, Cursor, Jules, …) | Drop the repo in (or beside) your project — the agent auto-reads AGENTS.md. |
| Claude Code | git clone … ~/.claude/skills/talisman, then run /talisman. |
| Cursor | Add a rule pointing at the repo, or reference SKILL.md in .cursor/rules/. |
| Windsurf | Reference SKILL.md from .windsurfrules. |
| GitHub Copilot | Reference SKILL.md from .github/copilot-instructions.md. |
| Aider | aider --read SKILL.md (or --read AGENTS.md). |
| Anything else | Tell your agent: "Follow the Talisman workflow in SKILL.md." |
Then just describe the deliverable — "make a branded PDF guide of how this app works" — and the agent follows the seven-phase workflow.
You: build an internal guide for this codebase, in our brand style
Agent:
1. extracts your brand palette, logo, and fonts from the app
2. reads the data model and features to get the content right
3. scaffolds docs/guide/ and writes the chapters + TikZ diagrams
4. builds the branded cover + subtle page background
5. compiles with pdflatex and renders pages to verify visually
→ docs/guide/main.pdf
A seven-phase workflow (full detail in SKILL.md):
| 1 | Extract brand identity Palette from CSS ( oklch() → hex), logo, font character. |
| 2 | Research real content Data model, lifecycle, feature modules, integration points. TOC by necessity. |
| 3 | Scaffold Copy the templates, fill the brand placeholders. |
| 4 | Write chapters Grounded in the code, with worked setup + handling use-cases. |
| 5 | Add diagrams Domain model, a flow, a state machine (TikZ). |
| 6 | Cover + background Full-bleed brand cover with the logo; faint watermark on content pages. |
| 7 | Build & verify visually Compile, then render pages to PNG and actually look. |
talisman/
├── AGENTS.md # cross-agent entry point (the AGENTS.md standard)
├── SKILL.md # same workflow in Claude Code's skill format
├── icon.png # skill icon
├── scripts/
│ ├── oklch_to_hex.py # CSS oklch() brand colors → LaTeX hex
│ └── prepare_logo.py # webp/png logo → PNG + sample its background color
├── assets/
│ ├── preamble.tex # branded core: fonts, palette, boxes, header, background
│ ├── main.tex # full-bleed cover + document shell
│ └── Makefile # 3-pass pdflatex build
└── references/
├── environment.md # toolchain probing, engine/font fallbacks, error playbook
└── diagrams.md # TikZ recipes for the 3 diagram families + anti-collision rules
The skill probes the environment and adapts, but the happy path wants:
| Need | Tool |
|---|---|
| PDF engine | pdflatex (default; lualatex/xelatex only if their font runtime works) |
| LaTeX packages | tikz, eso-pic, tcolorbox, hyperref, listings, booktabs |
| Fonts (fallback) | helvet, Avant Garde (pag) — or the closest installed match |
| Color / logo conversion | python3 + Pillow (webp/png); rsvg-convert/inkscape for svg |
| Visual verification | pdftoppm (poppler) |
references/environment.md documents the fallbacks when any of these are missing.
oklch_to_hex.py — convert modern CSS colors LaTeX can't readpython3 scripts/oklch_to_hex.py 0.26624 0.15944 267.227
# oklch(0.26624 0.15944 267.227) -> #090E71 rgb(9, 14, 113)
prepare_logo.py — .webp logo → PNG + background colorpython3 scripts/prepare_logo.py public/logo.webp figures/logo.png
# prints the logo's background color for the cover color-match trick
Beyond the happy path, the skill captures the messy realities of building LaTeX on an arbitrary machine, so future runs don't relearn them:
pdflatex-first — many installs ship a broken lualatex font runtime..webp/.svg logo conversion, and the cover trick of matching the cover fill
to the logo's square so a monogram floats seamlessly.step, state), overfull
lines from long code paths, multi-pass builds for cross-references.Issues and PRs welcome — new diagram recipes, font-fallback mappings, and environment fixes are especially useful.
MIT © atexy4ba
Design enforcement with memory — keeps your UI consistent across a project
Create animation-rich HTML presentations using a coding agent's frontend skills
1000+ skills curated from Anthropic, Vercel, Stripe, and other engineering teams
Agent harness performance optimization with skills, instincts, memory, and security