Paper Summary Agent

Purpose

Use this skill whenever the user asks to summarize, review, analyze, compare, or organize an academic paper, preprint, technical report, or benchmark paper.

Required behavior

1. Accept any of the following inputs:
   • paper title
   • DOI
   • arXiv / OpenReview / CVF / ACM / IEEE / Springer / Semantic Scholar / project page URL
   • direct PDF URL
2. Use live web search when the paper identity or canonical source is uncertain.
3. Prefer the best available source in this order:
   • direct PDF URL
   • open-access paper page with PDF
   • canonical abstract/paper page
   • metadata-only page
4. Save fetched source files in the configured download directory from scripts/paths.py (default: ~/Documents/paper-summary-agent/downloaded/).
5. When a PDF source is available, extract embedded figures into <summary_dir>/<slugified_title>/, where <summary_dir> is the configured Obsidian summary directory from scripts/paths.py, recover full-width / two-column figures as cleanly as possible, and embed them in the final Markdown so the figures render directly in the report.
6. When a PDF source is available, also extract major tables and render them as Markdown tables inside the final report instead of leaving them as prose-only notes.
7. Use subagents explicitly for non-trivial paper tasks:
   • resolver for identity resolution and source gathering
   • method-analyst for architecture/method details
   • experiment-analyst for datasets, metrics, results, and limitations
   • figure-analyst for Korean explanations of extracted figures
   • table-analyst for Korean explanations of extracted tables and metric interpretation
   • writer for final Korean Markdown synthesis
8. Final output file must be a Korean Markdown file saved to the configured Obsidian summary directory (PAPER_SUMMARY_AGENT_SUMMARY_DIR; default: ~/Desktop/obsidian/summary_paper/). Do not save final reports under ~/Documents/paper-summary-agent/summary/; that directory is not the final Obsidian summary location.
9. Agent-facing instructions, filenames, code comments, and internal templates should remain in English unless Korean is necessary in the final report.
10. If full text cannot be obtained, still produce the summary and clearly mark Full text unavailable near the top.
11. Avoid unsupported claims. Distinguish observed facts from inference.
12. Keep long verbatim quotes to a minimum.
13. By default, all paper-analysis subagents must use the same current-session model family and reasoning profile as the parent agent. For example, if the current session is running gpt-5.5 with high reasoning / fast mode, subagents should use gpt-5.5 with the same reasoning profile. Do not pin or prefer an older model such as gpt-5.4; omit explicit model overrides when the runtime can inherit the parent model, unless the user explicitly asks for a faster, cheaper, or different option.
14. Default to a thorough report rather than a brief summary. Do not stop at a high-level abstract-style overview when the full text is available.
15. The experiment analysis must explicitly cover:

• evaluation targets: what is predicted / classified / retrieved / segmented, the unit of evaluation, and the task setting or split
• evaluation metrics, how each metric is defined or computed, and how higher/lower values should be interpreted
• evaluation protocol and training/inference setup
• main quantitative results
• ablations, efficiency measurements, and limitations
• figure-by-figure and table-by-table notes for the main paper, plus important appendix tables/figures when relevant

16. Every extracted figure used in the report must have a Korean explanation based on the paper's content, and those explanations should be drafted from the figure-analyst subagent output. Do not use the raw English caption as the main figure description.
17. Every extracted table rendered in the report must be accompanied by a Korean explanation and metric-reading guidance based on the table-analyst subagent output.
18. Insert automatically generated topic / venue tags near the metadata block in the final Markdown.
19. If the first analysis pass is missing evaluation targets, metric definitions, protocol details, or important figure/table explanations, do a second pass before writing the final report.

Workflow

Step 1: Interpret input

Classify the input as one of:

• title
• doi
• url
• pdf_url

If ambiguous, search the web before proceeding.

Step 2: Resolve canonical source

Run:

• python scripts/resolve_paper.py "<INPUT>"

This produces candidate metadata. If the result is weak or incomplete, improve it with live web search.

Step 3: Fetch source

When a usable URL exists, run:

• python scripts/fetch_paper.py "<URL>"

Then extract text:

• python scripts/extract_text.py <FETCHED_PATH>
• python scripts/normalize_text.py <TXT_PATH>

Then build a role-scoped context manifest:

• python scripts/context_router.py <TXT_PATH>

Use the generated .context.json as the default subagent input map. Do not fan out the full normalized paper text to every subagent unless a role-specific context is missing or clearly insufficient.

Step 4: Delegate to subagents

For substantial paper analysis, explicitly create subagents:

Model policy: spawn every paper-analysis role with the current parent/session model and reasoning profile. Prefer inheritance by omitting explicit model overrides when the runtime supports it; set overrides only when needed to match the active session or when the user explicitly requests a different model. Never hard-code an older default such as gpt-5.4.

Context policy: pass each role only its relevant entry from .context.json plus the small metadata, figure, or table manifests it needs. The writer should receive structured outputs from the analysis roles and request targeted snippets only for unresolved claims. This keeps method, experiment, figure, table, and writer prompts from repeatedly carrying the whole paper.

• resolver
  • confirm canonical title
  • list source URLs
  • note whether PDF/full text was obtained
• method-analyst
  • use the current parent/session model and reasoning profile; prefer inherited model settings over explicit version pinning
  • identify problem, main idea, model/pipeline, latent-vs-language mechanics, training/inference details
  • explain the role of key method figures and any probing setup used for interpretation
• experiment-analyst
  • use the current parent/session model and reasoning profile; prefer inherited model settings over explicit version pinning
  • identify datasets, evaluation targets, metrics, metric definitions, evaluation protocol, baselines, main tables/figures, appendix evidence, ablations, efficiency, failure modes, and limitations
• figure-analyst
  • use the current parent/session model and reasoning profile; prefer inherited model settings over explicit version pinning
  • write Korean explanations for each extracted figure
  • explicitly explain what the reader should focus on, how to read the figure, and why it matters
• table-analyst
  • use the current parent/session model and reasoning profile; prefer inherited model settings over explicit version pinning
  • write Korean explanations for each extracted table
  • explain what each metric measures, how to read the rows / columns, and the main takeaway
• writer
  • use the current parent/session model and reasoning profile; prefer inherited model settings over explicit version pinning
  • synthesize the final Korean Markdown file using the required template
  • integrate the outputs from figure-analyst and table-analyst
  • for each extracted figure, write 2-4 Korean sentences that explain what the figure shows, what the reader should focus on, and why it matters; these sentences must be placed directly below the figure image in the final report, not only in section 8
  • for each major extracted table, include a Markdown table in the final report rather than only prose references
  • for each major metric, explain what it measures, how to read it, and whether higher or lower is better when that matters
  • prefer detail and specificity over brevity when the source text supports it

Step 5: Save final Markdown

Write the final report to:

• <summary_dir>/YYYYMMDD_<slugified_title>.md
• If a PDF was fetched, call python scripts/save_summary.py "<title>" "<temp_markdown_path>" "<pdf_path>" so extracted figures are saved under <summary_dir>/<slugified_title>/, each figure explanation is verified directly below its image, extracted tables are rendered as Markdown tables, automatic tags are inserted near the metadata block, and section 8 serves as a compact figure/table index without duplicating full figure explanations.

Use this section structure:

• Original title:
• Input:
• Canonical URL:
• Source URLs:
• Date generated:
• Full text availability:

1. One-paragraph summary

2. Basic paper information

3. Problem definition

4. Core idea

5. Method details

5.1 Overall pipeline

5.2 Main modules

5.3 Training / inference details

6. Experimental setup

6.1 Datasets and evaluation targets

6.2 Metrics

6.3 Baselines

7. Main results

8. Figure / table notes

9. Strengths

10. Limitations

11. Critical commentary

12. Follow-up research ideas

13. Confidence and limitations of this summary

Minimum expectations for section depth:

• 6.1 Datasets and evaluation targets must state what is evaluated, the prediction unit, the dataset split or scenario, and any special setting such as zero-shot / few-shot / cross-dataset evaluation.
• 6.2 Metrics must define the reported metrics, not just name them.
• 6.2 Metrics must explain how to interpret the metric values, including whether higher or lower is better and any thresholds / averaging choices that matter.
• 6.3 Baselines must describe what each baseline changes relative to the proposed method.
• 7. Main results must include the main quantitative comparisons, not only verbal conclusions.
• 8. Figure / table notes must cover each major figure/table in the main paper, and relevant appendix items when they materially affect interpretation.
• 8. Figure / table notes must explain figures and tables in Korean. Do not copy the raw English caption as the main figure description.
• Major tables in the main paper should appear as actual Markdown tables in the report, not only as prose descriptions.
• When extracted figures are available, place each figure image and its Korean explanation near the most relevant section whenever possible. The explanation must appear immediately below the corresponding image. Use section 8. Figure / table notes as a concise placement/index summary plus table-focused notes; do not duplicate long figure explanations there after the figures have already been placed inline.
• When extracted tables are available, place them near the most relevant results section whenever possible, and use section 8. Figure / table notes as a concise placement/index summary rather than the only location where tables appear.
• 13. Confidence and limitations of this summary must state what was directly observed versus inferred, and mention if any comparison numbers came from cited prior work rather than a reproduced run in the paper.

Recommended structure inside the report:

• In 6.1, explicitly cover evaluation target, prediction unit, dataset / split, and task setting.
• In 6.2, for each important metric, explain definition, interpretation, and comparison caveats when needed.
• In 8, use entries such as `Fig. 1` -> `5.2 Main modules` as a compact index when the figure is already embedded inline. Put the full Korean figure explanation directly below the image itself. For tables, include Korean notes in 8 when the table is not already accompanied by a nearby explanation.

Trigger hints

Use this skill automatically when requests include phrases such as:

• summarize this paper
• review this paper
• explain this paper
• compare these papers
• organize these papers into markdown
• read this arXiv paper
• paper title: ...
• DOI: ...

Important

If the user provides no detailed prompt beyond the paper input, still run the full workflow and produce the Markdown file automatically.