Skip to content

luminary19/skillscribe

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SkillScribe

Turn what you learn into a record you can actually see.


Stars License Tools Changelog

Tools  •  Changelog  •  License  •  Issues

🧰 Tools

SkillScribe currently works alongside these AI coding tools:

Claude Gemini

Tool Vendor Status
Claude Anthropic ✅ Supported
Gemini Google ✅ Supported

More integrations on the way — see the Changelog.

🛠️ How it works

SkillScribe watches YouTube videos with Gemini (natively, via URL — no downloading) and selectively mines the reusable techniques into Claude Code skills. Most of a talk is filler, ads, and intros; SkillScribe extracts only what's worth keeping, scores it, and prunes the rest.

The watch is two-pass: Gemini's one paid video call produces exhaustive timestamped notes (verbatim code/commands, quotes, reasoning — anchored by a free caption transcript fetched via yt-dlp), then the notes are structured into scored skills. In the agent flow Claude does that structuring for free; the standalone CLI uses a cheap Gemini text-only call instead.

request ─▶ discover ─▶ watch (Gemini) ─▶ structure ─▶ generate ─▶ triage + prune ─▶ harden + optimize
 "watch 5    (yt-dlp)    exhaustive notes:    notes →     score-     keep/revise/      3 rounds, skill-creator
  videos on              video+audio+free     skills      gated      discard +         guidelines: rewrite body
  prompting"             transcript; excludes (Claude,    per tech   Suggested Agents  + trigger-rich description
                         ads/intros           free)

Two layers:

  • Agent surface — a callable /skillscribe Claude Code skill plus a skillscribe-reviewer subagent (.claude/). This is what you invoke; it orchestrates the loop and confirms cost before watching, then runs the official skill-creator skill on each survivor to bring it up to Claude's skill standard.
  • Tool layer — the skill_miner Python package the agent drives, with deterministic subcommands.

🚀 Usage

As a Claude Code agent (recommended)

From the repo, just ask:

/skillscribe watch 5 videos on prompting released in 2025

The agent discovers the videos, shows them with a cost estimate, asks before watching (each video is a paid Gemini call), generates skills, triages them into skills/ with a skills/REVIEW.md, then refines each survivor — three rounds (default) of hardening the body and optimizing the description using skill-creator's authoring guidelines (no eval/A-B testing). That refinement runs on your Claude subscription — no extra Gemini cost.

Directly via the CLI

# one-time setup
python -m venv .venv
.venv\Scripts\python.exe -m pip install -r requirements.txt
$env:GEMINI_API_KEY = "<your-key>"        # environment only; never written to a file

# free steps (no API calls)
.venv\Scripts\python.exe -m skill_miner parse    "watch 5 videos on prompting released in 2025"
.venv\Scripts\python.exe -m skill_miner discover "watch 5 videos on prompting released in 2025"

# paid steps
.venv\Scripts\python.exe -m skill_miner run "watch 5 videos on claude code tips" --dry-run   # 1 video, writes nothing
.venv\Scripts\python.exe -m skill_miner run "watch 5 videos on claude code tips" --yes        # full batch

Subcommands: parse · discover · watch <url> (add --notes-only to stop after the paid notes pass — saved to skills/_run/notes/ — and structure them with a stronger model) · generate <watch.json> · run. Global flags: --model (default gemini-3.1-pro-preview) · --threshold (default 6) · --output-dir (default skills) · --media-resolution (default low).

⚠️ Each watched video is a paid Gemini API call. To keep costs predictable, video is sent at low media resolution by default (~3× fewer tokens than high — roughly $0.25 for a 15-min video vs. ~$1.20). Raise it only if you need it: --media-resolution high. --dry-run watches exactly one video and writes nothing; a real batch (more than one video) requires --yes. GEMINI_API_KEY is read from the environment only.

📦 Requirements

  • Python 3.11+ (developed on 3.14)
  • A GEMINI_API_KEY (Google AI Studio)
  • Dependencies: google-genai, yt-dlp, and PyYAML (used by the bundled skill-creator) — all pinned in requirements.txt
  • No plugin install required — the skill-creator skill is bundled into the repo (Step 5), so cloning is enough.

🧩 Bundled skill-creator

SkillScribe vendors Anthropic's official skill-creator skill into .claude/skills/skill-creator/ so the package is self-contained: Step 5 (refining each mined skill against skill-creator's authoring guidelines) works for anyone who clones the repo, with no separate plugin install. Claude Code auto-discovers it as the project skill skill-creator. It's copied unmodified under the Apache License 2.0 — see .claude/skills/skill-creator/LICENSE.txt and VENDORED.md in that folder for provenance and how to refresh it. © Anthropic.

About

No description, website, or topics provided.

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors