| name | literature-review |
| description | Write a literature review by iterating over papers one at a time, accumulating summaries and themes across iterations. |
Shared files: progress.txt (done/remaining), learning.txt (cross-paper insights), research_note.md (summaries+themes), synthesis.md (final synthesis).
Lead Agent (loop controller — no reading/writing of papers)
- Init (first run only): copy
./agents/skills/templates/{progress,learning,research_note,synthesis}.* into working dir if not exist. List all PDFs; populate progress.txt with unchecked entries.
- Spawn Abstract Extraction Sub-Agent per paper in parallel via Task tool. Wait for all.
- Spawn Planning Sub-Agent once. Wait.
- Spawn Reading Sub-Agent for next unreviewed paper. Wait. Repeat until all done.
- Spawn Synthesis Sub-Agent once. Then stop.
Sub-Agents (all spawned via Task tool, general/mode: subagent)
Abstract Extraction (parallelizable, once per paper):
- Run
uv run python3 tools/extract_pdf.py <input.pdf>.
- Append
filename.pdf: abstract text to abstracts.txt. Append only; do not read.
Planning (once):
- Read
abstracts.txt.
- Reorder
progress.txt entries foundational-first. Do not mark any paper done.
Reading (sequential, once per paper):
- Read
progress.txt → pick next unreviewed paper.
- Read
learning.txt for accumulated themes/gaps.
- Extract paper:
uv run python3 tools/extract_pdf.py <input.pdf>.
- Read
templates/research_note.md for format; append new entry to research_note.md. Append only; do not read research_note.md.
- Append to
learning.txt (format from templates/learning.txt). Note contradictions; avoid restating existing themes.
- Mark paper done in
progress.txt with one-line note.
Synthesis (once, after all papers done):
- Read
research_note.md and learning.txt.
- Read
templates/synthesis.md for format.
- Overwrite
synthesis.md with all three sections: Synthesis, Contradictions, Open Questions.
Writing style (all sub-agents):
- concise by sacrificing grammer but avoid ambiguity
- define jargon
- LaTeX for equations (display mode for key ones)
- max 3 sentences/paragraph
- paragraph-based narrative
- no bullets/tables
- conversational tone
Writing style (learning.txt):
- speak aloud your thought process as you read each paper.
- keep the reflection conncise and focused on insights/themes/gaps.
- if any issue in tool usage, note it here and move on.
Stop Conditions
- All papers reviewed +
synthesis.md written
- Sub-agent fails repeatedly → escalate to user
- User cancels