Wiki Synthesis Publishing

Wiki Synthesis Publishing

Publish a finished multi-source research investigation to the LLM-Wiki as a canonical synthesis page plus a findings manifest. This is the synthesis-publishing pattern — distinct from:

• Ingest (llm-wiki op 1): user hands ONE raw source → ingest it.
• batch-research-ingest-workflow.md (inside llm-wiki): "research X and add to wiki" — search → extract → ingest, usually one concept per source.
• This op: an investigation ALREADY done (multiple sources synthesized) → publish as ONE canonical concept page + ONE manifest + N immutable raw ingests.

NOTE on llm-wiki: The llm-wiki umbrella skill is the canonical home for wiki operations, but in this environment it is flagged created_by=None (manually authored) and is off-limits to autonomous curation. This skill captures the synthesis-publishing sub-pattern that llm-wiki does not yet document. If llm-wiki becomes editable later, this content could be merged into its Core Operations as "2b. Research Investigation → Canonical Page + Manifest" (the exact section text is preserved in references/llm-wiki-section-2b.md).

When this skill activates

• User: "turn this research into a wiki page / push a canonical page + manifest into the wiki"
• After a research session produced a synthesis the user wants persisted as a wiki page
• You've finished a multi-angle investigation and want it to compound in the wiki

Prerequisites

• A wiki at WIKI_PATH following the llm-wiki SCHEMA (SCHEMA.md, index.md, log.md, raw/ immutable).
• The research is DONE — you have sources, a synthesis, and (ideally) confidence calibration.

Procedure

① Orient (critical — do this every session):

• Verify WIKI_PATH exists. On macOS the env may say LLM-Wiki but the real dir is llm-wiki (HFS+ hides case; Linux breaks on the mismatch). ls -d "$WIKI_PATH" first.
• Read SCHEMA.md, index.md, recent log.md. Search for existing pages on the topic before creating anything (avoids duplicates / missed cross-links).

② Extend taxonomy FIRST if a new domain appears. If the topic has no tag section (e.g. creative-writing), add a new #### Topic Section to SCHEMA.md with needed tags before using them (SCHEMA rule: declare tags before use). One patch.

③ Ingest sources as raw files under raw/articles/<category>/:

• Write body + raw frontmatter (source_url, ingested: YYYY-MM-DD, sha256: TODO).
• Compute REAL sha256 over the whole file and backfill — never ship TODO/placeholder:

  shasum -a 256 "raw/articles/<cat>/<file>.md" | cut -d' ' -f1
  

  Then patch the sha256: line. Lint flags placeholder/fake hashes as [INFO] placeholder.
• Raw is IMMUTABLE. Never edit a raw file after ingest.

④ Write the findings manifest at raw/articles/<category>/findings-YYYY-MM-DD.md:

• Research question, sources investigated, key findings, contradictions/tensions (documented, NOT resolved), confidence table, deliverables.
• It IS a raw file → give it raw frontmatter and a REAL sha256 too (don't fake the manifest hash).

⑤ Write the canonical concept page at concepts/<slug>.md (NOT a raw file — full wiki frontmatter):

• type: concept, composes: [] (root synthesis — no composed_by backfill needed), contested: true when genuine tensions exist, workflow: developing.
• confidence: high ONLY when primary sources were ingested directly; medium for contested/cognitive claims.
• Body: synthesized playbook; append ^[raw/articles/<cat>/<file>.md] provenance markers to paragraphs tracing to a specific source; include a "## Related pages" section with ≥2 [[wikilinks]] to existing pages so it isn't isolated.

⑥ Update index.md — add the new concept entry (alphabetical), bump 'Total pages' and 'Last updated'. Batch all index changes into ONE patch call (sequential patches duplicate entries); use execute_code / Python for 3+ entries.

⑦ Append to log.md — one entry: agent, trigger, files created (concept + manifest + N raw ingests), taxonomy extension, index bump, confidence, contradictions noted. Prefer Python append over patch (avoids anchor/pipe-prefix corruption — see llm-wiki Refresh pitfall):

p = f"{WIKI}/log.md"
s = open(p).read()
entry = "\n## 2026-07-21 create | <subject>\n- Agent: Research ...\n- ...\n"
open(p,'w').write(s.rstrip()+'\n'+entry)

⑧ Report every file created/updated to the user.

Pitfalls

• Placeholder sha256 in the manifest — the manifest is a raw file; hash it for real.
• Taxonomy-before-use — tags used in a page but absent from SCHEMA.md are a lint error. Extend first.
• composed_by backfill not needed — a root synthesis page has composes: []; only pages that BUILD on it later get composed_by. Don't invent backlinks.
• WIKI_PATH case — env may say LLM-Wiki; real dir is llm-wiki. Orientation ① catches it.
• Batch index updates — one patch for all index additions; sequential patches duplicate entries.
• Log append via Python — never patch-append to log.md (anchor/pipe-prefix corruption).
• Protected llm-wiki — do not attempt to edit llm-wiki itself (created_by=None, off-limits to autonomous curation). This skill is the sanctioned satellite for the synthesis-publishing pattern.

References

• references/llm-wiki-section-2b.md — the exact "2b" section text to merge into llm-wiki if it ever becomes editable, plus the worked example from this session (creative-writing investigation → canonical page + 5 raw ingests + manifest).
• references/worked-example-creative-writing.md — full command sequence + file inventory from the novel-craft investigation that spawned this skill.