update

Update Skill

Update an existing processed document with new information, corrections, or re-verification against updated sources. Every update ends with a mandatory CLI-grounding pass on the changed content - no silent skip - because any content change can introduce an ungrounded claim.

Pre-flight install (MANDATORY - run every session, no asking)

python3 -c "import stellars_claude_code_plugins" 2>/dev/null || python3 -m pip install --user --upgrade stellars-claude-code-plugins

No-op when the package is importable; auto-installs when missing OR when a stale shim is on PATH but the package is uninstalled in the active Python. Never ask - just run the line.

When to use

• New source document added to 1-input/ - re-run affected sections
• User found errors in 3-output/ document - correct and re-verify
• Grounding audit failed - fix unconfirmed claims
• Uniformization rules changed - re-apply to existing output
• Additional context available - enrich existing document

Steps

1. Identify what exists: Read 3-output/ for the current document, 2-wip/ for the processing history, INSTRUCTIONS.md and BENCHMARK.md for the original program.

2. ASK user:

   • What changed? (new source, found error, rule change, enrichment)
   • Which output document to update?
   • Full re-run or targeted update?

3. Targeted update (default - faster):

   • Read the existing output document
   • Apply the specific change (new source integration, error correction, rule reapplication)
   • Re-run uniformization on affected sections
   • Update the document in place (versioned backup to 2-wip/<task-name>/<filename>_prev.md first)

4. Full re-run (if structural changes needed):

   • Update INSTRUCTIONS.md with new context
   • Re-execute from Phase 1 of the process skill using existing + new sources
   • Produces new version in 3-output/

5. MANDATORY grounding pass (every update, no exception): run the grounding CLI on the updated content via the grounding skill -

   • document-processing extract-claims --document <updated doc> --output validation/claims.json, review validation/claims.json
   • document-processing batch-ground --claims validation/claims.json --source <src1> --source <src2> ... --output validation/grounding-report.md (one --source per 1-input/ source; pass --semantic on — the default — unless the user has opted out, i.e. settings.semantic_enabled == false; see the grounding skill)
   • document-processing check-consistency --document <updated doc> --output validation/consistency-report.md Fix any UNCONFIRMED / CONTRADICTED claim and any consistency finding before declaring done. Apply the grounding skill's verdict rules.

6. Re-verify against benchmark: Run BENCHMARK.md evaluation against the updated document. Report the score delta.

Rules

• Always create a backup of the current output before modifying: copy to 2-wip/<task-name>/<filename>_prev.md
• New source documents go to 1-input/ first - never process from arbitrary locations
• Grounding is re-run on any content change - this is not optional and not "if time permits"; step 5 is a gate
• 1-input/ stays read-only; never modify source material
• Update the WIP manifest with what changed, when, and the post-update grounding score
• The grounding mechanics live in the grounding skill - this skill calls them, never re-implements them