AMindToThink/consolidate-paper

Consolidate Research into a Living Paper

Create a single living markdown document that consolidates scattered research artifacts (notes, logs, configs, experiment outputs, submodule docs) into a structured research paper.

Output path: $ARGUMENTS (default: docs/paper.md)

Phase 1: Inventory Sources

Before writing anything, gather and read all source material in parallel:

1. Research documents: drafts, writeups, preregistration, research logs, READMEs
2. Experiment configs: YAML files, parameter settings
3. Results artifacts: metrics.json, stats.json, or equivalent results files
4. Plot sidecar files: .description.txt, .commentary.txt files that describe figures
5. Submodule docs: READMEs, PLAN.md, HYPOTHESES.md from submodules
6. Verify which figures exist: Glob for actual plot files (png/pdf) so you reference real files, not imagined ones

Use parallel tool calls aggressively — read all independent sources at once.

If any markdown files are too large (e.g., Google Docs exports with base64 images), use Grep pattern="^[^!]" to extract text lines only.

Phase 2: Plan the Structure

If the user provided a plan, follow it. If not, propose a standard structure and get approval:

• Title, Authors, Abstract
• Introduction (motivation, assumptions, contributions)
• Related Work (organized by topic, not flat list)
• Methods (one subsection per methodology component)
• Experiments & Results (one subsection per experiment, with tables and figures)
• Discussion (observations, implications, limitations)
• Conclusion
• References (consolidated, deduplicated)
• Appendices (pre-registration, supplementary analyses, hypotheses, etc.)

Phase 3: Write the Document

Write the full document in one Write call to avoid consistency issues from piecemeal assembly. Key principles:

• Embed figures with relative paths from the doc's directory: ![Caption](../outputs/experiment/plots/figure.png)
• Include full results tables — copy exact numbers from metrics/stats files, don't summarize
• Use tag conventions for incomplete sections:
  • <!-- TODO: description --> — work needed
  • <!-- NOTE: description --> — informal notes, links, planning items
  • <!-- FEEDBACK: question --> — questions for collaborators
• Don't invent data — if results aren't available, use a TODO tag
• Mark planned/proposed work with TODO tags inline rather than in a separate section — keeps them close to the relevant context

Phase 4: Verify

After writing:

1. Count TODO/NOTE/FEEDBACK tags: grep -c '<!-- TODO' docs/paper.md
2. Check line count is reasonable
3. Spot-check that results tables match source data files
4. Verify all referenced figures actually exist (Glob for each path)
5. Check for duplicate references

Phase 5: Companion README

Create a brief README in the same directory explaining:

• What each file is
• The tag conventions and how to find gaps
• The relationship between the paper and canonical source documents (logs, preregistration, etc.)
• How figures are referenced
• Conversion plan (e.g., pandoc to LaTeX when ready)

Common Pitfalls

• Large markdown files with embedded images: Use Grep, not Read. See the read-large-md skill.
• Stale line-number references: If following a plan from a prior conversation, line numbers may be wrong. Re-read files fresh.
• Trusting research logs over data files: If there's a discrepancy between a log entry and the actual metrics.json, the JSON is authoritative.
• Over-emphasizing one finding: Let the user decide what's interesting. Present results factually and let the framing come from discussion.