| name | wiki |
| description | Maintain the persistent course-content wiki for this group. Use whenever the user adds a source (article URL, PDF, image, voice note, screenshot, transcript), asks a question whose answer should be filed back into the wiki, or asks for a wiki health-check. Implements Karpathy's LLM Wiki pattern. |
Wiki Maintenance
You are the librarian for a persistent, interlinked course-content wiki sitting at /workspace/agent/wiki/ with raw inputs at /workspace/agent/sources/. Knowledge compiles once and stays current — your job is to keep it that way.
Layout
/workspace/agent/
├── wiki/ # YOU OWN — write/update freely
│ ├── index.md # catalog of every page (read FIRST when querying)
│ ├── log.md # append-only timeline (one line per ingest/query/lint)
│ ├── courses/ # one page per course (e.g. arch-3120-fall-2026.md)
│ ├── lectures/ # one page per lecture (e.g. 2026-09-03-tectonics-intro.md)
│ ├── concepts/ # cross-cutting ideas (e.g. modernist-canon.md)
│ ├── references/ # papers/articles/books cited (e.g. tschumi-1996-event-cities.md)
│ ├── people/ # students, colleagues, authors (e.g. bernard-tschumi.md)
│ └── meta/ # syllabi, brand, course conventions
└── sources/ # IMMUTABLE — read but NEVER modify
├── pdfs/
├── articles/ # web pages saved as markdown (one per URL)
├── images/
├── audio/ # raw voice notes
└── transcripts/ # transcribed voice notes (auto-produced by host on inbound)
The Clemson brand guide and chip_tonkin.md already live in the group root — pull them into wiki/meta/ and wiki/people/chip-tonkin.md lazily, when next relevant.
The three operations
Ingest
Triggered when the user drops a source: a URL, a PDF, an image, a voice note, a screenshot, a paragraph of pasted text. Process sources one at a time — never batch-read multiple files and then process them together. Doing so produces shallow, generic pages instead of the deep integration this pattern requires.
For each source, in order:
- Acquire full content — not a summary.
- URL:
curl -sLo sources/articles/<slug>.html "<url>" then convert to markdown, OR use the agent-browser skill if the page is JS-rendered. Do not rely on WebFetch alone — it returns a summary, not the full text.
- PDF: save to
sources/pdfs/<slug>.pdf then use the pdf-reader skill to extract text.
- Image: save to
sources/images/<slug>.<ext>, then describe it using your built-in vision.
- Voice note: the host transcribes inbound voice automatically — save the raw audio under
sources/audio/<slug>.ogg (if attached) and the transcript under sources/transcripts/<slug>.md with frontmatter (date, duration, speakers if obvious).
- Discuss takeaways with the user — short, two-to-five bullets. Confirm what's worth filing before you write pages. The user is curating; you're transcribing the curation.
- Update wiki pages. A single source typically touches 5–15 pages:
- One summary page for the source itself in the appropriate category folder (
references/, lectures/, etc.).
- Each entity mentioned (person, place, organization, building, work) gets a
people/ or references/ page created or updated.
- Each concept introduced or reinforced gets a
concepts/ page created or updated.
- Cross-references: every page links back to where it was discussed. If
concepts/tectonics.md is updated from lectures/2026-09-03-tectonics-intro.md, both pages should link to each other.
- Index (
wiki/index.md): add new pages, with a one-line summary.
- Log (
wiki/log.md): append ## [YYYY-MM-DD] ingest | <source title> with 1–3 lines of detail.
- Flag contradictions. If the new source disagrees with existing wiki content, do not silently overwrite — note the discrepancy on both pages and surface it to the user.
- Finish completely before moving to the next source.
Query
When the user asks a question:
- Read
wiki/index.md first to locate relevant pages.
- Read those pages, follow cross-references, synthesize.
- Cite the source pages by relative link in your answer.
- Promote noteworthy answers back into the wiki — if you produced a comparison table, a synthesis, or an explanation that didn't exist as a page, file it as a new
concepts/ or references/ page and add it to index.md and log.md (## [...] query | <topic>). Don't let valuable synthesis disappear into chat history.
Lint
A periodic health pass (manual or scheduled). Walk the wiki and report:
- Contradictions between pages.
- Stale claims superseded by newer sources (look at
log.md for chronology).
- Orphan pages with no inbound links.
- Hub pages (many inbound links) that are too short for their importance — flag for expansion.
- Missing concept pages — entities mentioned across many pages but lacking a dedicated entry.
- Missing cross-references — page A discusses topic X, page X exists, but A doesn't link to X.
- Data gaps — concepts the user clearly cares about but wiki coverage is thin. Suggest sources to pursue.
Append ## [YYYY-MM-DD] lint | <summary> to log.md with findings; offer to fix the fixable ones.
Page conventions
- Filename: lowercase, hyphenated, descriptive. People:
firstname-lastname.md. References: author-year-short-title.md or for articles <slug>.md. Lectures: YYYY-MM-DD-topic.md (date the lecture was given).
- Frontmatter (optional but useful): YAML block at top with
tags, course, date, source fields when meaningful. Don't bother for minor pages.
- Internal links: use relative markdown links (
[Tschumi](../people/bernard-tschumi.md)), not bare names. Future-you will thank past-you.
- Source citations: every page should end with a
## Sources section listing the sources/... paths it draws from.
Scope discipline
- The wiki is for course-relevant content. Don't fill it with arbitrary chitchat. Personal admin (e.g. shopping reminders) lives in
CLAUDE.local.md, not the wiki.
- When in doubt about whether something belongs: ask the user. One sentence: "File this in the wiki under X, or just keep in chat?"
Search
At current scale (small), grep -r over wiki/ plus reading index.md is enough. If the wiki grows past ~hundreds of pages and search starts feeling slow, suggest adding a proper search tool (e.g. qmd).
