-
Notifications
You must be signed in to change notification settings - Fork 0
Everyday Operations
type: reference up: "What-is-the-llm-wiki-Template" related:
- "What-is-the-llm-wiki-Template"
- "Wiki-Structure-and-Conventions"
- "Best-Practices"
- "Team-Workflow"
- "Knowledge-Graph" source: "https://github.com/crcresearch/llm-wiki-memory-template/tree/main/.claude/skills" tags: [procedure, operations, slash-commands]
What a user does day to day with a template wiki: ingest, query, lint, and let the agent commit. Grounded in the .claude/skills/wiki-*.md and .claude/commands/wiki-*.md files, the SCHEMA's Operations section, and the verification/write-protocol gate documents.
Tell the agent to file it, or run /wiki-source. The procedure (from .claude/skills/wiki-source.md): read the source; discuss the takeaways with you and confirm the framing and cross-link targets; create a type: source-summary page named after the source, with up: to the nearest parent, source: set to the URL or path, and any clear extends:/supports:/criticizes: edges; update related pages so the new source reinforces or revises them (and flag any contradiction rather than leaving it); fix cross-references both ways; update the index under "Source summaries"; append a log entry; optionally rebuild the graph; run the verification gate; then commit. A typical source ingest touches 5 to 15 pages.
Run /wiki-experiment or ask the agent to file the run. Per .claude/skills/wiki-experiment.md: read the existing pages for this experiment first (integrate, do not duplicate); create or update a results page (type: synthesis or entity, up: to the benchmark page, with typed edges where the result confirms or refutes a prior claim); record configuration, headline metrics, what changed versus the previous run, and what was surprising, plus links to the results/ directory and the commit hash. Honest reporting is mandatory: bad and contradicted results are filed plainly, and no number comes from a projection. Then update affected concept pages, fix cross-references, update the index and log, run the verification gate, and commit.
Read the index, drill into the relevant pages, and synthesize. The Query procedure: read index_<repo>.md to find candidate pages, read them, answer with citations to page names. For topology questions ("what connects to what") use the knowledge graph; for content questions ("what does this page say") read the files. If the synthesized answer is reusable, offer to file it back as a new page (often type: analysis) so the exploration compounds.
Run /wiki-lint or ask for a health check. Per .claude/skills/wiki-lint.md it scans for orphan pages, dead links, stale claims, missing frontmatter, type: untyped pages, missing concept pages, missing cross-references, index gaps, naming deviations, and special-file integrity. It reports findings grouped by type, asks which to fix this pass (lint is incremental), applies the accepted fixes with bidirectional cross-reference repair, updates the index, appends a log entry, and commits. See Best-Practices for cadence.
Three, each a user-invokable front door to a model-side skill:
| Command | Skill | Use it to |
|---|---|---|
/wiki-source |
wiki-source.md |
Ingest an external source (paper, article, README, design doc) |
/wiki-experiment |
wiki-experiment.md |
File experiment results as durable memory |
/wiki-lint |
wiki-lint.md |
Health-check the wiki and apply fixes |
The slash commands are a safety net for explicit invocation; the proactive default is that the agent runs these operations itself when the situation calls for it, per the wiki-maintenance rules in CLAUDE.md.
It runs the source-ingest procedure above. Invoke it when a new external document should enter the project's memory, and you want to force the ingest explicitly rather than rely on the agent's proactive behavior. It is not for your own experiment results; use /wiki-experiment for those.
It runs the experiment-ingest procedure above. Invoke it after a run produces metrics, ablations, or comparisons worth keeping. Skip it for a debug rerun with no new information, or a run whose configuration is identical to one already filed.
It runs the lint procedure above. Invoke it every several sessions, after a major ingest that touched many pages, or when you notice a broken link or a contradiction. It is incremental, so you need not fix everything at once.
You can, though the pattern is LLM-authored-by-default. If you edit by hand, follow the SCHEMA: valid frontmatter with type: and up:, an opening summary line, a See also section, and cross-references in both directions. Then let the agent (or yourself) run the verification gate and commit. Hand edits are fine; they just should meet the same conventions the agent's writes do.
Tell it what changed and let it integrate. The template's stance is that new information triggers an ingest: the agent updates the affected page, revises anything the change contradicts (flagging rather than silently overwriting), fixes cross-references, and logs it. If the new information contradicts a wiki claim, the rule is to trust what is observed now and update or flag the stale page, not repeat it.
The Verification Gate (wiki/agents/verification-gate.md) is the pre-commit checklist that every wiki write runs after saving and before git add/git commit. It re-reads each changed page from disk and checks: numerical claims backed by real outputs and tagged with corpus scope; no cross-corpus gap arithmetic; every new typed edge paired with a body-level back-reference; body links resolve; frontmatter valid with type:/up:; analysis/decision pages have their required sections; index and log updated; Home updated for category-level changes; and honest reporting throughout. If any criterion fails, fix it and re-run; do not commit until all pass. It fires on every ingest, experiment, and lint write.
The wiki-write-protocol (wiki/agents/wiki-write-protocol.md, backed by scripts/wiki-write-protocol/protocol.sh) governs pushing the wiki sub-repo when multiple agents may write concurrently. It fires in place of a plain git push, typically after an ingest or lint when you ask to publish. It does an optimistic push with retry: index/log conflicts union-merge mechanically via .gitattributes, and genuine content conflicts are left with markers for the agent's next turn to resolve. Note: the reference implementation ships and is CI-tested, but it is not yet wired into the overlay's default write path, so today you invoke it deliberately when pushing a shared wiki. See Team-Workflow.
- What-is-the-llm-wiki-Template — orientation
- Wiki-Structure-and-Conventions — the files these operations touch
- Best-Practices — cadence, honest reporting, and what to file
- Team-Workflow — pushing a shared wiki safely
- Knowledge-Graph — rebuilding the graph as part of ingest