Paper Micro-Unit Generator

Use this skill only for local repair or an explicitly requested legacy scaffold. paper-formal-writer remains the sole formal author and the only producer of paper_output/final_paper_source.md and paper_output/final_paper.docx.

Entry Check

For formal repair, run:

python .claude/skills/paper-workflow-orchestrator/scripts/workflow_guard.py --skill paper-micro-unit-generator

Continue only when S0-S6 pass and paper_output/qa/repair_queue.json contains at least one item whose strategy is micro-repair. Do not create a repair queue manually to bypass the two-failure threshold.

Legacy and quickstart modes are installation or old-model scaffolds. They do not satisfy S7 and do not require the formal repair queue.

Formal Repair Mode

1. Read paper_output/plan/writing_plan.json, paper_output/context/authoring_state.json, paper_output/qa/draft_audit.json, and paper_output/qa/repair_queue.json.
2. Select only queued micro-repair items. Keep the issue ID, section, failure category, expected evidence, and attempt count intact.
3. Read the affected section draft and only the evidence needed by that issue. Do not rewrite unrelated sections.
4. Write repair material to paper_output/drafts/repairs/<issue-id>.md. The material must contain concrete replacement text, its target location, and any required evidence marker.
5. Apply the replacement to the queued section draft. Preserve valid headings, equations, figure/table identifiers, citations, and evidence markers.
6. Re-run the section audit:

python .claude/skills/paper-formal-writer/scripts/validate_authoring.py --section <section-id>

7. Return to $paper-formal-writer or $paper-workflow-orchestrator. This skill must not assemble the formal manuscript or generate Word.

If the same blocking category reaches the third changed attempt, respect authoring_state.status = BLOCKED. Report the evidence or requirement causing the failure and suggest Lite as a user decision; never switch editions automatically.

Repair Rules

• Repair the failed claim or paragraph, not the whole paper.
• Use only computed values and paths present in the current evidence chain.
• Preserve <!-- mathmodel-evidence: ... --> markers until deterministic assembly removes them.
• Do not invent data, citations, completed experiments, or successful validation.
• Do not expose workflow language in paper prose.
• Do not renumber existing figure, table, equation, or bibliography references.
• A repair is complete only after validate_authoring.py --section returns PASS.

Legacy Mode

The old command names remain available. With no arguments they write only to paper_output/drafts/legacy/:

python .claude/skills/paper-micro-unit-generator/scripts/generate_all_offline.py
python .claude/skills/paper-micro-unit-generator/scripts/merge.py

Expected outputs:

paper_output/drafts/legacy/micro_units/*.txt
paper_output/drafts/legacy/generate_log.json
paper_output/drafts/legacy/legacy_scaffold.md
paper_output/drafts/legacy/legacy_scaffold.docx
paper_output/drafts/legacy/legacy_ref_check.md

tasks.json is required only for this legacy batch path. The merge is deterministic and preserves existing reference numbers.

For the built-in smoke test, the orchestrator passes --output-root paper_output/quickstart --stem quickstart_scaffold; all smoke-test files stay under paper_output/quickstart/.

Never write final_paper.md, final_paper_source.md, final_paper.docx, or similarly formal names from legacy/quickstart mode.

Template Library

Read references/micro-unit-library.md only when a queued repair needs a fine-grained paragraph/sentence pattern or when the user explicitly requests legacy scaffolding. Do not load the full 200+ template library for ordinary section writing.

Handoff

After a formal repair, run workflow status and update persistent context:

python .claude/skills/paper-workflow-orchestrator/scripts/workflow_guard.py --status
python .claude/skills/context-memory-keeper/scripts/update_workflow_memory.py

The guard report is authoritative if memory and current artifacts disagree. Read the persistent snapshot at paper_output/context/workflow_memory.json before resuming a repair.