inspire

inspire — a Claude Code plugin

Drop a batch of links into Claude Code — YouTube videos, web articles, or a mix — and get each one evaluated against the project you're working in: what it's about, what you can learn from it, and what's directly applicable here — synthesized into a small durable corpus under docs/inspiration/. Then, when you want to act on it, /apply promotes the vetted leads into your project's own docs.

It works in two stages. Intake (/inspire):

• A skill, /inspire — the orchestrator. Collects the URLs you drop, routes each to the right reader, fans them out, then writes the notes + a synthesized index.
• Two subagents — inspire-watcher evaluates one YouTube video, inspire-reader evaluates one web page. Both are read-only and minimal-egress by design (see Security); each returns a finished evaluation in the same shape.
• An MCP server, inspire-content — a thin, pure-read content fetcher the subagents call. Two tools (get_youtube_transcript, get_webpage_content) in a self-contained single Python file with PEP 723 inline dependencies.

Promotion (/apply):

• A skill, /apply — the editor. Reads the vetted corpus and proposes a doc-by-doc edit brief, then applies only what you approve, per edit.
• A subagent, inspire-applier, and a second MCP server, inspire-docs — the applier executes approved edits through the server's one tool, write_doc, which is path-bounded to your docs directory. That makes "only edits project docs" a structural fact, not a promise (see Security).

Requirements

• Claude Code with plugin support.
• uv on your PATH — the MCP server is launched via uv run, which provisions its Python deps in an isolated, cached environment on first use. Nothing is installed into your project.

Install

/plugin marketplace add skqr/claude-inspire-plugin
/plugin install inspire@inspire-marketplace

Then, in any repo:

/inspire https://youtu.be/VIDEO_ONE
https://www.youtube.com/watch?v=VIDEO_TWO
https://example.com/some-article

URLs can be one-per-line or mixed into a sentence — every YouTube and web URL in your message is picked up (YouTube links go to inspire-watcher, other web pages to inspire-reader). Results land in docs/inspiration/ (one note per source + an index with a cross-cutting-themes synthesis).

Then, when you want to turn those leads into changes:

/apply

/apply reads the docs/inspiration/ corpus, presents a doc-by-doc edit brief (which file, what change, why), and applies only the edits you approve — one at a time. Its writes are confined to your docs directory (INSPIRE_DOCS_DIR_PATH, default ./docs); you can also point it at a specific note or theme (/apply the agentic-engineering note).

Security model

Fetched content — a video transcript or a web page's text — is untrusted third-party content and a genuine indirect-prompt-injection surface: captions, titles, or page body can contain text engineered to hijack a reading agent. This plugin is built around that:

1. Least-privilege readers. Each reader is allowlisted to exactly one content tool plus read-only repo access (Read/Grep/Glob). No Bash, no Write/Edit, no general web. A successful injection has no shell to run and no file to mutate.
2. The orchestrator writes files, not the exposed agent. Readers return their evaluation as text; the /inspire skill writes docs/inspiration/. The write capability never sits in a context that has ingested raw fetched content.
3. Data-not-instructions framing. Content is delivered inside fences with a standing instruction to treat it as data; injection-looking text is reported, never obeyed.
4. The MCP server is a pure read. It only resolves a URL → text. It strips terminal/control characters and caps length so a source can't pipe escape sequences or a wall of text into a model's context.