ai-skills

A personal collection of skills for Claude Code (and any agent with a tool loop — they are used through Cursor and Hermes too).

The skills differ in kind. Some collect fresh content. Some drive other tools or other models. What they share is the split: a stateless script does the mechanical work, and the agent writes the result. The script never formats prose, and the agent never scrapes or polls by hand.

Each skill is a folder under skills/ with its own SKILL.md and scripts/. Copy the folder you want into your agent's skills directory. Nothing else in this repo is needed at runtime: a skill that shares code carries its own copy, and tests/test_standalone_skill.py runs every skill from a copied folder to keep that promise true.

Skills

Installation

Install what the skill you copied needs.

The collectors need Python packages:

pip install -r requirements.txt

The consensus panel needs no Python and no API key. It needs jq, the claude CLI, and the Cursor CLI (agent login). It runs both on their own subscription.

Adding a new skill

1. Create the folder skills/<name>/. Put a SKILL.md and a scripts/ directory in it.
2. Give SKILL.md YAML frontmatter with name and description. The name must equal the folder name, in lowercase and hyphens. The test suite checks this.
3. Keep the script stateless. Let it print structured material. Let the agent write the prose.
4. Use no machine-specific paths. Take configuration from config.yaml, from flags, or from environment variables.
5. Run the tests. Then commit and push. The repo is built to grow as a collection.

If the new skill collects content and needs fuzzy dedup, take shared/dedup.py — but copy it into the skill's own scripts/ and import it from there. shared/ is the source, not a runtime dependency: a skill that imports it from the repository root stops working the moment its folder is copied out. The test suite compares each copy with the source byte for byte, so edit shared/dedup.py and copy it over the others.

Content collectors

┌─────────────────────────────────────────────┐
 │ Agent (LLM) layer                           │
 │ (scheduled runs, persistence, formatting)   │
 └─────────────────────────────────────────────┘
                  ↓
 ┌─────────────────────────────────────────────┐
 │ Stateless collector scripts                 │
 │  ai-daily ⇄ RSS        trending-skills      │
 │  (filter + RapidFuzz dedup, print metadata) │
 └─────────────────────────────────────────────┘
                  ↓
 ┌─────────────────────────────────────────────┐
 │ External sources (RSS, web, GitHub API)     │
 └─────────────────────────────────────────────┘

Each collector is stateless and idempotent: it reads a small SQLite "seen" database, marks what it printed, and on the next run returns only genuinely new items. Stdout carries two blocks (СТАТЬИ metadata and СОДЕРЖАНИЕ content) that the LLM turns into the final post.

Why dedup lives in the script

A collector that does exact-hash dedup only lets near-duplicates through: the same incident under a reworded title. These scripts use RapidFuzz partial_ratio (threshold ~65) plus the exact hash, with a token-overlap fallback when RapidFuzz is absent. So "OpenAI launched GPT-5" and "OpenAI releases GPT-5, all details" collapse into one item.

Model panel

consensus sends one question to models from different vendors. Round one collects independent answers. Round two shows each model the other answers with the names removed, and asks it to revise. The script prints the material and stops there; the agent writes the synthesis and keeps the disagreements, because agreement between models is not proof.

Both harnesses run read-only. Claude runs without Edit, Write and Bash. Cursor runs with --mode ask.

Tests

python -m pytest tests/ -v

The suite checks the pure functions of the collectors, the dedup helper, and the frontmatter of every SKILL.md against the Agent Skills spec. CI runs it on Python 3.11, 3.12 and 3.13.

License

MIT