← ClaudeAtlas

deep-reviewlisted

Expert review workflow for papers, PDFs, arXiv links, markdown drafts, and technical artifacts. Use when the user wants a rigorous review that interviews them first, researches claims, checks math/methods/literature, uses specialist passes when useful, and returns severity-ranked findings with evidence.
dipakkrishnan/deep-review-skill · ★ 0 · Code & Development · score 72
Install: claude install-skill dipakkrishnan/deep-review-skill
# Deep Review Use this skill to review papers and other technical artifacts with the depth of a serious referee, domain expert, or adversarial collaborator. ## Start Here 1. Identify the artifact source: local PDF, arXiv URL, markdown/text draft, repository artifact, or pasted content. 2. Check for review memory: user-level in `~/.deep-review/` and project-level in `.deep-review/`. If present, read it before the calibration interview and summarize any preferences you plan to apply. 3. If the artifact is a local PDF or arXiv URL and text extraction is needed, prefer the scripts in `scripts/`. 4. Before doing substantive review, read `references/orchestrator_protocol.md`. 5. Ask a short calibration interview before reviewing. Use `references/calibration_questions.md` to select questions. 6. Choose the review depth: - `quick`: at least 2 first-wave specialist passes, 1 critique round. - `standard`: at least 4 first-wave specialist passes, 1 critique round. - `deep`: at least 8 first-wave specialist passes, then a defense/rebuttal round and an independent adjudication round. Tiers are first-wave floors, not caps; expand with follow-up passes when findings warrant. 7. Produce the final report using `references/output_format.md` and the taxonomy in `references/review_rubric.md`. 8. At the end, ask whether to save or update review memory. Only write to `~/.deep-review/` (profile, preferences) or `.deep-review/` (recurring issues, artifact history) after explicit user