Paper PPT Deep Research
Overview
Use this skill after reading agent_task.json and the extracted paper assets. It organizes the uploaded paper into focused research passes so the final deck is faithful, detailed, and coherent.
Workflow
- Read
source_assets/paper.md and source_assets/figures.json or figures.md.
- Create
research/deep/plan.md with focused passes. Typical passes:
- problem, motivation, and research gap
- method/architecture and algorithm details
- data, experiments, metrics, baselines, and ablations
- figures/tables/equations worth showing
- limitations, assumptions, failure modes, and implications
- slide narrative and audience framing
- If the runtime provides Task/SubAgent tools, assign focused readers when deep research is enabled. Use separate readers for the paper's background/related work, method, experiments, and critique when the paper is complex. Skip only when the tool is unavailable, fails, or the paper is too short/simple for meaningful decomposition; record the concrete reason in
agent_report.json.subagents.
- Store pass notes under
research/deep/notes/. Keep notes factual, with section/page/figure anchors where available.
- Run
scripts/compile_deep_notes.py to write research/deep/notes_index.json, even when a failed/unavailable SubAgent leaves the index empty and the limitation must be described.
- Write
research/deep/brief.md with slide-ready synthesis and conflicts/uncertainties.
When this skill is enabled, the backend blocks manuscript.md, design_spec.md, notes, agent_report.json, and slide SVG authoring until research/deep/notes_index.json and research/deep/brief.md exist.
Merge the synthesis into manuscript.md; do not paste independent reader styles into the deck.
Script
Use the Python interpreter from agent_task.json.paths.python or PAPER_PPT_PYTHON.
"<python>" skills/paper-ppt-deep-research/scripts/compile_deep_notes.py \
--notes-dir research/deep/notes \
--out research/deep/notes_index.json
The script only indexes notes you or SubAgents already wrote. It must not replace paper reading or decide what matters.
Quality Rules
- Tie findings back to the uploaded paper. Prefer concrete paper facts, metrics, figures, tables, equations, and named components.
- Separate what the paper proves from your interpretation or external context.
- Record SubAgent usage or skip reasons in
agent_report.json.subagents.
- Keep the final deck narrative unified; the main Agent owns synthesis and style consistency.
1---2name: paper-ppt-deep-research3description: Deep paper analysis for Paper PPT Agent decks. Use when Agent mode has deep research enabled or when a long/technical paper needs focused reading passes, SubAgent/task decomposition, evidence extraction, limitation analysis, and a slide-ready synthesis before manuscript and SVG generation.4---56# Paper PPT Deep Research78## Overview910Use this skill after reading `agent_task.json` and the extracted paper assets. It organizes the uploaded paper into focused research passes so the final deck is faithful, detailed, and coherent.1112## Workflow13141. Read `source_assets/paper.md` and `source_assets/figures.json` or `figures.md`.152. Create `research/deep/plan.md` with focused passes. Typical passes:16 - problem, motivation, and research gap17 - method/architecture and algorithm details18 - data, experiments, metrics, baselines, and ablations19 - figures/tables/equations worth showing20 - limitations, assumptions, failure modes, and implications21 - slide narrative and audience framing223. If the runtime provides Task/SubAgent tools, assign focused readers when deep research is enabled. Use separate readers for the paper's background/related work, method, experiments, and critique when the paper is complex. Skip only when the tool is unavailable, fails, or the paper is too short/simple for meaningful decomposition; record the concrete reason in `agent_report.json.subagents`.234. Store pass notes under `research/deep/notes/`. Keep notes factual, with section/page/figure anchors where available.245. Run `scripts/compile_deep_notes.py` to write `research/deep/notes_index.json`, even when a failed/unavailable SubAgent leaves the index empty and the limitation must be described.256. Write `research/deep/brief.md` with slide-ready synthesis and conflicts/uncertainties.2627When this skill is enabled, the backend blocks `manuscript.md`, `design_spec.md`, notes, `agent_report.json`, and slide SVG authoring until `research/deep/notes_index.json` and `research/deep/brief.md` exist.28Merge the synthesis into `manuscript.md`; do not paste independent reader styles into the deck.2930## Script3132Use the Python interpreter from `agent_task.json.paths.python` or `PAPER_PPT_PYTHON`.3334```bash35"<python>" skills/paper-ppt-deep-research/scripts/compile_deep_notes.py \36 --notes-dir research/deep/notes \37 --out research/deep/notes_index.json38```3940The script only indexes notes you or SubAgents already wrote. It must not replace paper reading or decide what matters.4142## Quality Rules4344- Tie findings back to the uploaded paper. Prefer concrete paper facts, metrics, figures, tables, equations, and named components.45- Separate what the paper proves from your interpretation or external context.46- Record SubAgent usage or skip reasons in `agent_report.json.subagents`.47- Keep the final deck narrative unified; the main Agent owns synthesis and style consistency.
crui5in