devon-claude-skills

Turn a web article into a faithful-capture literature note in an Obsidian vault. Use when the user wants to ingest a web page (blog post, essay, technical article, documentation page) into their personal source notes corpus. Fetches the page, extracts structured content (TL;DR, takeaways, faithful body prose, references). The article sibling of youtube-synthesizer — same vault conventions, same frontmatter base schema, no frame pipeline.

Manage the PR review feedback loop: monitor CI checks, fetch review comments, and iterate on fixes. Use when: (1) pushing changes to a PR and waiting for CI/reviews, (2) user says "new reviews available", (3) iterating on PR feedback from Gemini, Cursor, Claude, or other reviewers, (4) monitoring PR status. Supports multiple review bots: Gemini Code Assist, Cursor Bugbot, and Claude agent fallback. Also supports custom agent reviewers defined in AGENT-REVIEWERS.md for focused reviews (security, DRY, etc.). Automatically detects priority levels from different bot formats and handles rate limits.

Mine this project's Claude Code session logs into durable improvements — proposed memories, CLAUDE.md rules, docs, skills, and decision-log (DECISIONS.md) entries. Async "AI dreaming": distill raw sessions into friction-focused digests, then consolidate across sessions (dedup vs what's already captured, including any project DECISIONS.md, frequency-gated promotion). Use when: (1) the user asks to mine/learn-from past sessions, "dream over my logs", or improve project memory/docs from experience; (2) a nightly/periodic self-improvement pass; (3) after a stretch of work, to capture recurring lessons. NOT for reviewer-roster tuning — that's the `dream-reviewers` skill.

Validate the agent reviewers that gate a project's quality (the pr-review-loop roster). Mine every PR's review threads + the dream-marker stream into per-reviewer scorecards (acceptance rate, false-positive/won't-fix/reopen rates, taste overrides), then propose roster changes (keep / refine / retire). Use when: (1) the user asks to validate/audit/tune their PR reviewers or AGENT-REVIEWERS.md, (2) deciding whether a reviewer earns its keep or should be retired/refined, (3) a periodic reviewer-health pass. For mining general session lessons, use the `dream` skill instead.

Compact the current conversation into a handoff document so a fresh agent (new session, different machine, or a teammate's chat) can pick up the in-flight work without ramping from zero. Use when: (1) ending a long session with state that won't survive compaction or memory alone, (2) handing work to a teammate or a different agent (Claude Code → Codex, Claude → Cursor), (3) parking work for later when the to-resume context is too large to reconstruct from git history + beads alone.

Turn a YouTube video into a faithful-capture literature note in an Obsidian vault, with optional Dropbox sync via rclone. Use when the user wants to ingest a video (explainer, howto, lecture) into their personal source notes corpus. Combines transcript + load-bearing visual frames + structured extraction. Auto-discovers vaults under ~/ObsidianVaults/ and offers to push the new entry to dropbox: after writing.

Download a YouTube video and extract frames at specified timestamps with perceptual hashes; also discovers candidate timestamps via ffmpeg scene-detect + per-second pHash run-grouping; analyses optical flow over a sub-range and composes a sprite strip showing a motion arc. Use when capturing specific moments from a video as PNG images, or when a downstream skill (e.g. youtube-synthesizer) needs a high-recall list of where things change in the video. Mechanical only — no LLM calls.

Extract transcripts from YouTube videos. Use when the user asks for a transcript, subtitles, or captions and provides a YouTube URL or 11-character video ID. Supports plain text, timestamped output, language selection, and writing directly to a file.

Generate, edit, and compose images using Google's Gemini 3 Pro Image model (Nano Banana Pro). Use this skill when the user asks to create images, generate visuals, edit photos, compose multiple images, create logos, thumbnails, infographics, product shots, or any image generation task. Supports text-to-image, image editing, multi-image composition (up to 14 images), iterative refinement, aspect ratio control, and Google Search-grounded image generation for real-time data visualization.