Public-Agent

Self-contained agent skills, rules and hooks for coding agents. Skills and rules are
agent-independent: a skill is a Markdown contract (SKILL.md) plus plain scripts, a rule is
a plain Markdown file, so any agent that reads Markdown and runs shell commands can use them.
Hooks depend on each agent's hook protocol; the ones here speak three. Every folder under
skills/ works on its own: no shared code, no external config, nothing installed outside the
folder.

This repository is a small public part of my own agentic harness: the skills, rules and hooks
that work without the rest of it, stripped of every link to my private tools and projects.
AGENTS.md serves as the master index and operating agreement across all subfolders.

Skills

Skill	What it does	Needs
`blast-radius`	Maps reverse imports, export signature changes, broken call sites and affected tests for files or a `git diff`	Node.js, `git` (for `--diff`)
`mediagrab`	Downloads one public video as mp4 or its audio as mp3 via yt-dlp	Windows, Node.js
`prompt-expansion`	Expands a brief task into a structured XML prompt; trigger `+++`	—
`transcribe`	Turns a local file or a YouTube URL into `M:SS <text>` lines, via captions or Groq Whisper	Windows, Python 3.13, Groq API key
`webcopy`	Saves one web page with all its assets as an offline copy	Python 3, `requests`, `beautifulsoup4`

Rules

Rule	What it sets
`reply-format.md`	Wording, skeleton and grounding markers of every agent reply
`ticket-template.md`	How a work request becomes a ticket: format, vocabulary, lifecycle

Rules carry no frontmatter. Copy rules/ next to the agent's instruction file (AGENTS.md,
CLAUDE.md or the equivalent of the agent in use) and link or include the rules from there.
They reference each other and the prompt-expansion skill, so copy them together with it.

Hooks

Hook	What it does	Agents
`pii-guard.js`	Replaces e-mails, phone numbers, IBANs and public IPs in tool output with `[PII:…]` placeholders, blocks prompts that contain them, denies reads of `.env` files and credentials without asking, optionally asks before reads of dumps, databases, uploads or logs	Claude Code (full), Antigravity and Qwen Code (partial)
`quota-guard.js`	Reads context size and cumulative session tokens from the transcript before the model runs and blocks prompts that exceed the model class threshold at 0 token cost until unlocked via `!go`, `!force` or `allow`	Antigravity, Claude Code, Qwen Code

Quick start

# 1. Install: copy a skill folder into the agent's skill directory (<skill-root>)
Copy-Item -Recurse skills/mediagrab <skill-root>/
# 2. Binaries (mediagrab, transcribe): each skill runs its own setup into its own bin/
powershell -ExecutionPolicy Bypass -File <skill-root>/mediagrab/scripts/setup.ps1    # ~118 MB, network yes, no cost
powershell -ExecutionPolicy Bypass -File <skill-root>/transcribe/scripts/setup.ps1   # ~220 MB, network yes, no cost
# 3. Key (transcribe only): enter GROQ_API_KEY in config.json beside SKILL.md
Copy-Item <skill-root>/transcribe/config.example.json <skill-root>/transcribe/config.json

Agents with native skill loading find a skill by the name and description in its
frontmatter. Claude Code, for example, reads ~/.claude/skills/ as <skill-root>.
Any other agent gets the path to SKILL.md in its instruction file, or in the prompt.

Notes

Binaries are not in the repository. scripts/setup.ps1 of mediagrab and transcribe
downloads yt-dlp.exe and ffmpeg.exe from their official releases and checks every SHA-256
against scripts/setup.lock.json. ffmpeg.exe alone exceeds GitHub's 100 MB file limit.
Windows only for mediagrab and transcribe: the scripts call the .exe builds in bin/.
transcribe uploads audio to Groq whenever no original YouTube captions exist. Private
recordings need the owner's consent first.
mediagrab handles public content only. Platform terms of service may forbid downloading;
that call stays with the user.

License

MIT

Public-Agent-Kit