WebinarNotes/CLAUDE.md

LLM Wiki Schema (Operational Contract)

You are the dedicated maintainer of this vault as a persistent, compounding wiki.

Mission

Maintain a high-signal personal knowledge base where:

• raw/ contains immutable source materials.
• wiki/ contains LLM-authored, continuously maintained markdown pages.
• index.md is the content catalog.
• log.md is an append-only chronological operation log.

The user curates sources and asks questions. You perform all wiki maintenance.

Non-Negotiable Rules

1. Never modify files inside raw/.
2. Always update index.md and log.md after any ingest/query/lint operation that changes the wiki.
3. Prefer editing existing pages over creating duplicates.
4. Use Obsidian-style wiki links ([[page-name]]) for internal references.
5. Record contradictions and superseded claims explicitly (do not silently overwrite history).
6. Keep claims attributable: include source links to wiki/sources/* pages.
7. Do not leave orphan pages intentionally; add at least one inbound and one outbound link when possible.
8. Keep writing concise, structured, and diff-friendly.
9. Every wiki page carries exactly one page-type hashtag (see Tagging Rules) matching its folder.

Folder Convention

raw/
  sources/            # immutable source markdown/text/pdf exports
  assets/             # immutable local images/files referenced by raw sources

wiki/
  overview.md         # top-level synthesis and navigation
  script-coverage.md  # machine-maintained: concept coverage vs the webinar script
  sources/            # one summary page per ingested raw source
  entities/           # people, orgs, projects, places, tools
  concepts/           # themes, ideas, methods, frameworks
  timelines/          # optional chronological reconstructions
  comparisons/        # side-by-side analyses
  queries/            # durable outputs created from Q&A sessions
  lint-reports/       # periodic health-check reports

File Naming Rules

• Use kebab-case for file names.
• Prefix source summary pages with date: YYYY-MM-DD-title.md.
• Prefer stable canonical pages:
  • wiki/entities/<name>.md
  • wiki/concepts/<concept>.md
• If a page name collides, merge instead of creating suffixes unless truly distinct.

Tagging Rules

Every page in wiki/ starts with its page-type hashtag on its own line, directly after the # <title> heading (Obsidian inline tag format):

# <Title>

#<type-tag>

## <first section>

Folder → required tag:

Folder / file	Tag
wiki/overview.md	#overview
wiki/script-coverage.md	#coverage
wiki/sources/*	#source
wiki/entities/*	#entity
wiki/concepts/*	#concept
wiki/timelines/*	#timeline
wiki/comparisons/*	#comparison
wiki/queries/*	#query
wiki/lint-reports/*	#lint-report

• Exactly one page-type tag per page; it must match the page's folder.
• Additional topical tags are allowed on the same line after the type tag (e.g. #concept #ai-tooling), but the type tag comes first.
• When creating any new page, add the tag immediately; when moving a page between folders, update the tag in the same operation.

Required Page Templates

1) Source Summary (wiki/sources/*.md)

Must include sections:

• # <title>
• #source tag line (per Tagging Rules)
• ## Source Metadata (date, raw path, source type, ingestion date)
• ## Core Claims
• ## Key Evidence / Details
• ## Connections (links to entities/concepts/comparisons)
• ## Open Questions
• ## Change Impact on Wiki (what pages were updated and why)

2) Entity / Concept Page

Must include:

• # <name>
• #entity or #concept tag line (per Tagging Rules)
• ## Summary
• ## Current Understanding
• ## Evidence (bullets with links to wiki/sources/*)
• ## Related Pages
• ## Contradictions / Uncertainty
• ## Next Questions

3) Query Output (wiki/queries/*.md)

Must include:

• #query tag line (per Tagging Rules)
• Question asked
• Answer
• Evidence trail (links)
• Follow-up questions
• Whether this output changed existing pages

4) Coverage File (wiki/script-coverage.md)

Machine-read by the dashboard plugin, which parses the table positionally. The shape below is required exactly — a deviation renders as parse errors, not as a best-effort read.

• # Script coverage
• #coverage tag line
• ## Metadata, containing these two lines verbatim in this form:
  • - **Script:** followed by the raw script path in backticks
  • - **Last synced:** followed by a YYYY-MM-DD date
• ## Coverage, containing one markdown table with exactly these four columns in this order: Concept, Status, Station, Pinned
  • Concept — an Obsidian wikilink to a page in wiki/concepts/
  • Status — exactly one of covered, partial, absent, lowercase
  • Station — one of the seven station names spelled exactly as listed in Workflow D, a comma-separated list of them, the lowercase word all, or an em dash — when the status is absent
  • Pinned — the lowercase word yes, or empty
• ## Notes and ## Related Pages are free prose and are not parsed.

The parser reads every markdown table in the file, so do not add a second table.

Standard Workflows

Workflow A: Ingest One Source

When user says "ingest ":

1. Read the raw source from raw/sources/ (and raw/assets/ references if needed).
2. Extract key claims, facts, entities, concepts, uncertainty.
3. Create/update one source summary in wiki/sources/.
4. Update relevant wiki/entities/* and wiki/concepts/* pages.
5. Update wiki/overview.md synthesis if this source materially changes understanding.
6. Update index.md.
7. Append ingest entry to log.md.
8. Report what changed, what is uncertain, and suggested next source/questions.
9. If this ingest created or modified any page in wiki/concepts/, run Workflow D (Sync Script Coverage) before reporting.

Workflow B: Answer Query

1. Read index.md first.
2. Select relevant pages and synthesize answer with page citations.
3. If answer has durable value, save to wiki/queries/YYYY-MM-DD-<slug>.md.
4. If answer reveals new synthesis, update affected concept/entity pages.
5. Update index.md and append query entry in log.md when files changed.

Workflow C: Lint Wiki

Run periodic health checks for:

• contradiction detection across pages
• stale claims superseded by newer sources
• orphan pages / weak linking
• high-mention concepts lacking dedicated pages
• missing evidence links
• missing or folder-mismatched page-type tags (per Tagging Rules)

Write report to wiki/lint-reports/YYYY-MM-DD-lint.md, then update index.md and log.md.

Workflow D: Sync Script Coverage

Maintains wiki/script-coverage.md — one row per page in wiki/concepts/, judged against the script named in that file's **Script:** metadata field.

1. Read wiki/script-coverage.md and note every row where Pinned is yes.
2. Read the script and every page in wiki/concepts/.
3. For each concept, decide Status and Station:
   • covered — the script delivers the idea, whether or not it uses the page's name.
   • partial — the script gestures at it but never lands it.
   • absent — the script never reaches it.
   • Station is one of the seven — Chat box, ReAct, Tools, Memory, Skills, Process, OS — spelled exactly as written here, or a comma-separated list of them, or the lowercase word all, or — when absent.
4. Never modify a row whose Pinned is yes — not its status, not its station. A pinned row is the user's judgment and outranks yours.
5. Add rows for concept pages with no row. Remove rows whose concept page no longer exists. A pinned row whose concept page no longer exists is still removed — step 4 protects the user's judgement about a live concept, not a dangling row.
6. Update **Last synced:** to today.
7. Update index.md and append a sync entry to log.md.

Run this workflow:

• at the end of any ingest that creates or modifies a page in wiki/concepts/
• whenever the script file itself changes
• on the explicit sync script coverage intent

The coverage baseline is always the raw script. Ingesting the script into wiki/sources/ does not change the baseline.

Citation & Evidence Policy

• Prefer citing wiki source summaries (wiki/sources/*.md) rather than raw files in normal answers.
• If citing raw material directly, also reflect it into a source summary page.
• Mark uncertain or contested claims with Status: tentative in relevant section.

Update Policy

• Never perform silent large rewrites.
• Preserve meaningful prior interpretations; move outdated material under a "Superseded" note when needed.
• Keep sections ordered consistently for predictable diffs.

Operational Commands (Natural Language)

Supported intents:

• "ingest path-or-title"
• "query: question"
• "lint wiki"
• "show recent changes"
• "suggest next sources"
• "sync script coverage"

Always execute intents according to workflows above.

Session Start Checklist

At start of every session:

1. Read index.md and the latest section of log.md.
2. Identify last completed operation and current open questions.
3. Continue from prior state without resetting conventions.