Demiurge is an AI coding skill that combines five disciplines into one workflow: lazy problem-solving (ponytail), terse output (caveman), human-sounding code (humanizer), frontend design sense, and strict coding standards. It works with TypeScript, JavaScript, C, C++, Python, Rust, Go, Java, Kotlin, Swift, Dart, HTML, CSS, and more.
It auto-detects what you want to do from natural language. Say what you need and it figures out the right approach.
npx skills add bthavanish/demiurge/plugin marketplace add bthavanish/demiurge
/plugin install demiurge@demiurgeClone this repo and copy skill/ into your agent's skills directory.
/demiurge [mode] [target]
No mode needed. Type what you want:
/demiurge # full audit + ask which fixes to apply
/demiurge fix the security issues in my api # context-gather for auth, fix it
/demiurge clean up this frontend code # context-gather, humanize it
/demiurge make a settings page # build it
/demiurge what can I delete from this repo # debt modeGive an exact task and it context-gathers for that change. No context and it runs a full audit.
UI guard: UI references only load for apps with a UI. Backend/CLI/library apps skip UI modes automatically.
| Mode | Command | Description |
|---|---|---|
make |
/demiurge make [feature] |
Builds on user instructions using ponytail ladder + coding standards. |
design-material |
/demiurge design-material [target] |
Material Design 3 builder. Explicit only, UI apps only. |
| Mode | Command | Description |
|---|---|---|
audit |
/demiurge audit [target] |
Full codebase audit. Security, logic, dead code, style, architecture. |
audit-frontend |
/demiurge audit-frontend [target] |
Frontend/UI/UX audit. Design tokens, a11y, responsive, anti-patterns. UI apps only. |
audit-backend |
/demiurge audit-backend [target] |
Backend/logic/security audit. Injection, auth, error handling, type safety. |
critique |
/demiurge critique [target] |
UX design review. Nielsen heuristics (/40), cognitive load, personas. UI apps only. |
review |
/demiurge review [target] |
One-line code review. Severity-tagged: bug, risk, nit, q. |
| Mode | Command | Description |
|---|---|---|
secure-code |
/demiurge secure-code [target] |
Fixes bugs, vulnerabilities, logic errors, dead code directly. |
humanize |
/demiurge humanize [target] |
Detects AI patterns in code and rewrites to sound human. |
polish |
/demiurge polish [target] |
Final quality pass. Typography, spacing, color, motion, copy. UI apps only. |
harden |
/demiurge harden [target] |
Production-readiness. Error handling, i18n, edge cases. |
bolder |
/demiurge bolder [target] |
Amplifies bland designs. More decisive and committed. UI apps only. |
quieter |
/demiurge quieter [target] |
Tones down aggressive or overstimulating designs. UI apps only. |
| Mode | Command | Description |
|---|---|---|
debt |
/demiurge debt [target] |
Harvests ponytail: comment markers into a debt ledger. |
compress |
/demiurge compress [filepath] |
Compresses natural language files to save tokens. |
commit |
/demiurge commit |
Generates terse conventional commit messages. |
| Mode | Command | Description |
|---|---|---|
docs |
/demiurge docs [target] |
Writes or restructures documentation using Diataxis framework. |
readme |
/demiurge readme [target] |
Creates or improves README files with audience-specific templates. |
prose |
/demiurge prose [target] |
Edits machine-generated prose to remove AI writing tropes. |
Every mode applies these rules:
- Ponytail ladder -- simplest solution first. YAGNI. Reuse before building.
- Caveman output -- terse output, full accuracy. Code blocks preserved verbatim.
- Humanizer -- no AI slop in code or prose. No em dashes, no rule-of-three, no AI vocabulary.
- Coding standards -- KISS, SRP, descriptive names, no magic numbers, immutability, portable, no hallucinations.
- Comment standards -- why not what, no dead code, no metadata, ponytail markers for shortcuts.
- DESIGN.md enforcement -- follow the design system if one exists.
- Modularity -- small, cohesive, loosely coupled functions. Dependencies flow one direction.
- No hardcoding -- env vars, config files, constants, parameters.
- Always ask before fixing. Never auto-fix without presenting findings and getting user confirmation.
Demiurge works with multiple AI coding agents through adapter directories:
| Agent | Adapter Directory |
|---|---|
| Claude Code | .claude/skills/demiurge/ |
| OpenCode | .opencode/skills/demiurge/ |
| Cursor | .cursor/skills/demiurge/ |
| Gemini | .gemini/skills/demiurge/ |
| Kiro | .kiro/skills/demiurge/ |
| Trae | .trae/skills/demiurge/ |
| Trae CN | .trae-cn/skills/demiurge/ |
| Pi | .pi/skills/demiurge/ |
| Qoder | .qoder/skills/demiurge/ |
| Rovo Dev | .rovodev/skills/demiurge/ |
| Generic | .agents/skills/demiurge/ |
| Codex | .codex/skills/demiurge/ |
Built on work from the security and developer tools community.
| Skill | Author | Repository |
|---|---|---|
| caveman | Julius Brussee | JuliusBrussee/caveman |
| humanizer | blader | blader/humanizer |
| impeccable | pbakaus | pbakaus/impeccable |
| ponytail | Dietrich Gebert | DietrichGebert/ponytail |
| material-3-skill | hamen | hamen/material-3-skill |
| maintainable-typescript & write-good-docs | miguelspizza | miguelspizza/skills |
40 skills from the Trail of Bits skills collection. Security audit methodologies, vulnerability patterns, and analysis techniques.
| Source | What We Took |
|---|---|
| Trail of Bits Testing Handbook | Security testing methodologies, tool configurations, vulnerability detection techniques |
| Wikipedia: Signs of AI Writing | 33 AI writing patterns for the humanizer skill |
MIT. Copyright 2026 Thavanish brijesh.
Third-party references are used under their respective licenses. See individual repositories for details. Most Trail of Bits skills use the MIT License.