Public skills maintained by Flint.
This repo is a simple home for reusable agent skills that are useful outside any one client or internal project. The goal is to keep each skill practical, inspectable, and easy to copy or install.
- A public skills repo for Flint-maintained agent workflows
- A place to publish skills that are broadly useful
- Complementary to larger shared/community skill repos, not a replacement for them
| Skill | Path | What it does |
|---|---|---|
analyze-video |
skills/analyze-video |
Analyze local videos or YouTube videos with a multimodal model — frame/UI/slide inspection, production critique, timestamped scene breakdowns. Gemini recommended; OpenAI fallback via frame sampling + Whisper. |
audio-analysis |
skills/audio-analysis |
Actually listen to audio files or YouTube audio — music/production feedback, track comparison, mystery-sound ID, Suno style-prompt extraction. Gemini recommended; OpenAI audio-input fallback. |
codex-lane |
skills/codex-lane |
Run OpenAI's Codex CLI (GPT-5.6) as a first-class execution lane inside a Claude Code session: a lane-split doctrine (what to delegate vs. keep), per-task reasoning-effort routing, the session pause/ask/resume protocol (codex exec resume), a verification contract, and Ubuntu 24.04 bwrap sandbox fixes. Claude orchestrates and reviews; Codex executes well-spec'd work. |
context-window |
skills/context-window |
Generate a single-file HTML report visualizing what's in an LLM session's context window — system prompt, tool schemas, recall, hooks, conversation turns — color-coded by source type with token estimates. Works for the calling agent's own session, for bots whose source you can read, and best-effort for black-box bots. |
de-ai-design |
skills/de-ai-design |
Audit an AI-generated web design for the convergent "AI look" (indigo gradients, pill badges, emoji UI, hover-lift cards, scroll reveals — the "Purple Problem") and replace each tell with a deliberate alternative from a chosen design genre. Evidence-based: every tell needs file:line proof in the actual CSS/JS/markup before it gets fixed. Ships a grep-able tells catalog. |
lyric-video |
skills/lyric-video |
Combine an audio file, a video file, and lyrics into a subtitled music video. Accepts plain-text lyrics (auto-aligned via whisper word timestamps) or pre-timestamped LRC / TSV / JSON. Two-pass build (composite + faststart remux) so a crash in either step doesn't destroy the work. |
people-registry |
skills/people-registry |
Set up a people.json entity-resolution system for agents — aliases + disambiguation hints, optional deep person context files, and the identity-verification rules that keep agents from treating display names as proof of identity. Scaffold via scripts/setup.sh; ships templates, a paste-ready ops snippet, and a fictional two-Sams worked example. |
read |
skills/read |
Fetch any web URL or YouTube video, cache it, and return content at three depth modes (skim / read / deep). Plus lightweight RSS/Atom feed subscriptions — subscribe, refresh, read by URL. Model-agnostic — the skill never calls an LLM, the calling agent does the synthesis. Caches to ~/data/read-cache/, subscriptions at ~/data/feeds/. |
safe-gdocs |
skills/safe-gdocs |
Read-only Google Docs / Drive access for agents. Wraps the official gws CLI with a guard that blocks every write method (create, update, delete, send, etc.) and ships a friendly gdocs read/search/list/info wrapper. Idempotent first-run installer handles npm install, shim placement, PATH check, and OAuth. |
site-archive |
skills/site-archive |
Archives a site or URL into markdown while respecting robots.txt, randomizing delays, supporting incremental crawls, and detecting blocker pages |
stock-research |
skills/stock-research |
End-to-end stock analysis for long-term investors. Live fundamentals/quote/earnings/analyst data via yfinance (no API key), public technical-analysis chart URLs from StockCharts + Finviz (no image downloads, no auth), and a four-question quality+valuation framework that lands a buy/watch/avoid verdict. Includes a second "Mainstreet" bottom-up mode (unit economics → discounted fair value → margin-of-safety entry) for monopoly / novel-business-model companies where P/E benchmarks mislead. |
suno-prompt |
skills/suno-prompt |
Turn a song idea into Suno-ready lyrics + a style prompt (under 1000 chars, no artist names) and settings. Forces a context/POV phase first so outputs don't collapse into generic AI mush. Chat-only and coding-agent paths; pairs with write-lyrics. |
wp-release-watch |
skills/wp-release-watch |
Hourly security tripwire for WordPress core releases: polls git tags, diffs each new release, runs an LLM security analysis over the patch + optional web threat intel, writes a markdown report, notifies via Slack webhook or any command. Model- and harness-agnostic. Idempotent setup.sh installs the cron. |
vuln-scan |
skills/vuln-scan |
Per-file CTF-style vulnerability scanner: loops source files through claude -p, writes *.vuln.md sidecars, and runs a skeptical verify pass to weed out false positives. Profiles for WordPress, Node, and Python. |
walkie |
skills/walkie |
Connect an AI agent to another agent over a direct P2P channel (walkie-sh / Hyperswarm DHT) — no server, no accounts, just a shared channel name + secret. Covers install (idempotent setup.sh), identity, connecting to a peer, the conversation-hygiene rules that keep two agents from blabbering forever (terminal sign-off token, spiral detection, banter caps), human-visibility patterns (watch/relay + the web UI), and per-peer security norms. Ships with zero real secrets — you generate your own. |
wordpress-alt-text |
skills/wordpress-alt-text |
Generate and apply screen-reader-first alt text to any WordPress site's images. SQLite-backed pipeline with inventory, AI-vision generation, review artifact, and staged write-back. |
wp-screenshots |
skills/wp-screenshots |
Capture clean WordPress admin + front-end screenshots from a JSON brief. Headless Chromium, login-aware, hides update bubbles, 2× DPR default, optional Mac-faithful font aliasing, standalone HTML gallery output. |
write-lyrics |
skills/write-lyrics |
Phased lyric-writing engine: messy draft → ruthless seed selection → structure + rhyme lock → canon iteration → observed AI-cliché filter. Chat-only or workspace files; optional scan-cliches.py. Pairs with suno-prompt. |
ytpoop |
skills/ytpoop |
Generate a short YouTube Poop-style chaotic absurdist video entirely programmatically — PIL frames, NumPy synth audio, ffmpeg assembly. No external assets. Ships with a runnable reference generator and a documented technique catalog for an agent to fork per topic. |
See CATALOG.md for the short index.
.
├── README.md
├── CATALOG.md
├── CONVENTIONS.md
└── skills/
└── <skill-name>/
├── SKILL.md # frontmatter + usage docs (required)
├── scripts/ # entry points + setup.sh (if any setup is needed)
├── examples/ # smoke-test inputs (optional)
└── README.md # only if SKILL.md isn't enough (optional)
Each skill lives in its own folder under skills/. See CONVENTIONS.md for the install-receipt pattern and the required setup.sh contract for skills that need setup.
Most skills are folder-based and use SKILL.md as the entry point.
Coding agents (Claude Code, Codex, Cursor, etc.) — point the agent at the skill folder, or clone this repo and load skills/<name>/SKILL.md.
skills/write-lyrics/
skills/suno-prompt/
ChatGPT / Claude.ai (no tools) — open SKILL.md on GitHub (or paste it into the chat) and tell the model to follow it. Process skills like write-lyrics and suno-prompt are written to run chat-only: they include labeled output blocks and don't require a filesystem. Reference files linked from SKILL.md help when the agent can fetch the raw GitHub URL; for a paste-only session, paste SKILL.md plus any references/*.md you care about.
Raw GitHub URLs (swap branch/path as needed):
https://raw.githubusercontent.com/flintfromthebasement/skills/main/skills/write-lyrics/SKILL.md
https://raw.githubusercontent.com/flintfromthebasement/skills/main/skills/suno-prompt/SKILL.md
For now this repo is maintained directly by Flint. Future public skills will be added here when they are stable enough to be useful outside local project work.
New skills follow the conventions in CONVENTIONS.md: one entry point per skill (scripts/setup.sh), idempotent installers with an install receipt, diagnostic output an agent can act on. safe-gdocs is the reference implementation.