| name | research-paper-analysis |
| description | Deeply analyze one or a few papers and turn them into reusable research notes. Use when Codex needs to read a paper beyond surface summary, extract the problem, claims, method, evidence, limitations, and relevance to a research goal, compare a small set of papers directly, or produce an auditable critique grounded in what the paper actually says. |
Research Paper Analysis
Overview
Read one or a few papers deeply enough to extract what matters and why it should be trusted. Start by writing an analysis plan, verify the target papers, read to the depth required by the task, build a claim-and-evidence record, critique the work, then produce a reusable research note or comparison.
This skill is not a generic summarizer. The output should preserve structure, uncertainty, and evidence so the resulting note can be reused in later research rather than re-read from scratch.
Workflow
- Write the analysis plan before reading deeply.
- Record the target papers, research goal, desired output, comparison scope, and reading depth.
- Decide whether the task is understanding, critique, comparison, implementation relevance, or replication assessment.
- Read analysis-plan.md and use its template.
- Verify and lock the paper set.
- Normalize metadata for each target paper and make sure you are reading the intended artifact.
- Freeze the paper set before deep reading. If you add another paper later, mark it as a scope change.
- Read for structure, not just for gist.
- Read title, abstract, introduction, method, experiments, figures, tables, and conclusion in the order most useful for the task.
- Escalate depth where the argument depends on details such as training setup, evaluation protocol, ablations, or theoretical assumptions.
- Read reading-and-claim-extraction.md.
- Build a claim-and-evidence record while reading.
- Record each important claim and the evidence the paper provides for it.
- Distinguish between what the paper states, what the evidence actually supports, and what you infer beyond the text.
- Read evidence-and-limitations.md.
- Critique before concluding.
- Identify assumptions, missing controls, weak comparisons, narrow evaluation, or overclaimed conclusions.
- Relate the critique to the task instead of generating generic reviewer language.
- Read critique-and-comparison.md.
- Use sub-agents only when the paper set is small but separable.
- Keep analysis goals, note schema, and final synthesis with the main agent.
- Delegate per-paper reads or clearly separated comparison slices.
- Require structured outputs that the main agent can merge without reinterpretation.
- Read subagent-protocol.md.
- Report the paper as a reusable research object.
- Return the paper set, analysis goal, structured findings, critique, confidence notes, and next actions.
- If comparing multiple papers, make the comparison dimensions explicit instead of stacking independent summaries.
- Read reporting-schema.md.
Operating Rules
- Read enough of the paper to support the requested judgment. Do not critique experiments you did not inspect.
- Keep claims and evidence separate. A strong-sounding paper can still have weak support.
- Mark inferences explicitly when they go beyond the paper text.
- Distinguish
central finding, supporting detail, and open question.
- Prefer fewer, higher-signal observations over bloated paraphrase.
- If the evidence is too thin for a strong conclusion, say so directly.
Read These References When Needed
