relearn

Two modes — user selects:

Mode A — Extract facts from article

Fast path (bulk)

For known topics in the TOPICS list, run the script directly — no sub-agents or file reading required:

bash
1python3 .claude/skills/generate-questions/scripts/relearn.py [topicId ...]
2# Omit topic IDs to process all known topics

The script fetches each article, calls claude -p to extract facts, and writes them back to the sources file.

Manual path (targeted)

For a sourceId not in the TOPICS list, or when AI judgment is needed during extraction.

For multiple sourceIds: spawn one sub-agent per topic, in sequence. Wait for each to complete before spawning the next. Sub-agents write directly to files — never return JSON blobs to the main context.

Step 1 — Identify source title and existing facts

Read assets/questions/sources/{topicId}.json. Note the title and facts fields for {sourceId}.

Step 2 — Fetch article text and cache it

bash
1python3 .claude/skills/generate-questions/scripts/fetch_wiki.py "{title}" \
2  | tee /tmp/wiki_{slug}.txt
3python3 .claude/skills/generate-questions/scripts/save_sources.py \
4  --topic {topicId} --source-id {sourceId} --article-text < /tmp/wiki_{slug}.txt

The tee output is the article text — use it for Steps 3–4. save_sources.py caches it to articleText (skips if already set). If fetch_wiki.py exits 3 (network down): abort, suggest Mode B with existing facts.

Step 3 — Extract new facts and save

From the article text, compose a JSON array of fact objects for each atomic fact not already in source.facts. Then pipe to save_sources.py:

bash
1echo '<json array>' | python3 .claude/skills/generate-questions/scripts/save_sources.py \
2  --topic {topicId} --source-id {sourceId} --facts

Fact object shape:

json
1{"id": "fact_{source_slug}_{NNN}", "text": "...", "verified": true, "verifiedAt": "{today ISO date}"}

Count {NNN} from the current max in source.facts. The script merges by id — new facts are appended, existing ones updated.

Step 4 — Re-verify existing facts

For each existing fact contradicted by the current article, pipe an updated array with those facts set to verified: false and a verificationNote:

bash
1echo '<json array of updated facts>' | python3 .claude/skills/generate-questions/scripts/save_sources.py \
2  --topic {topicId} --source-id {sourceId} --facts

Only include the facts that changed — the script merges by id.

Mode B — Cross-reference related sources

For a list of sourceIds (or a book set id):

Step 1 — Load all specified source entries

Read summary and facts only — do not load articleText (too large).

Step 2 — Compare across sources

Identify:

• Contradictions: fact A in source 1 contradicts fact B in source 2
• Corroborations: same fact confirmed by multiple sources
• Gaps: fact implied by one source but not recorded in another

Step 3 — Handle findings

Contradictions: mark both facts verified: false with a verificationNote citing the conflicting source.

Gaps: propose new facts to the user for review before writing.

Step 4 — Write updated sources files

For contradictions and approved gap facts, use save_sources.py --facts per source:

bash
1echo '<json array>' | python3 .claude/skills/generate-questions/scripts/save_sources.py \
2  --topic {topicId} --source-id {sourceId} --facts

Print cross-reference report: contradictions found, facts added, sources updated.

Modular inputs

Accepts sourceIds:

• As skill arguments (space-separated)
• From a book set file: assets/questions/book_sets/{superCategoryId}.json
• From stdin (pipe-friendly)

Token efficiency rules

1. Bulk runs → use relearn.py — not sub-agents
2. Multiple sourceIds (manual) → one sub-agent per topic, in sequence — write to files directly, never return JSON blobs; no parallel agents
3. Mode B → load only summary and facts, never articleText
4. Verify writes by exit code, not by re-reading the file