Agent Skills: end-to-end-study

end-to-end-study

Chain: best data for novelty -> low-hanging fruit -> most-likely paper (venue + story) -> author instructions -> method under rigor checklist -> reader/reviewer-oriented writing -> reviewer cycle -> private repo + tagged release.

Seven phases. Each phase links to a dedicated reference file.

Phase 0 - Orientation (30 s)

Confirm with the user: domain (oncology? immunology? neurology?), topic vagueness (is a topic given, or should the skill scan datasets first?), claim type preference (method / translational finding / benchmark). Decline phases outside scope (wet-lab validation, actual journal submission).

Phase 1 - Best data for novelty (data-first scan)

Novelty usually lives in under-mined assets of recently released large cohorts, not in re-analyses of classic datasets. Scan for (1) recent large-cohort releases with a secondary modality that has not been systematically exploited, (2) paired-modality data where one modality is under-used, (3) public drug-response or perturbation screens paired with rich clinical metadata.

See references/data-first-novelty.md for the scan heuristics, a catalog of currently-underutilised public datasets, and worked example (BeatAML ex-vivo drug-sensitivity table in Tyner 2018 supplementary).

Exit criterion: a primary dataset is named, its under-mined asset is identified, and a short list of 2-3 candidate claims is drafted.

Phase 2 - Low-hanging fruit selection

Within the chosen dataset, pick the claim with the best effort-to-impact ratio. See references/low-hanging-fruit.md for the selection matrix (combination-gap claim vs scale-gap vs rigor-gap vs orthogonal-outcome claim), with expected-IF mapping for each type.

Use WebSearch + bioRxiv / PubMed MCP to verify the chosen claim is actually un-done. Heuristics for keyword laddering and gap verification are in references/novelty-search.md.

Exit criterion: one sentence of the form "No prior work does X in Y; this paper shows Z (effect-size guess)", validated against literature.

Phase 3 - Target journal + author instructions

Pick the journal where this specific claim has the highest acceptance probability, not the highest IF in absolute terms. See references/journal-targets.md for the tiered catalog with acceptance-probability heuristics. Then immediately fetch the target journal's author instructions with WebFetch and distil them into the project-local manuscript/JOURNAL.md.

See references/author-instructions.md for the extraction checklist (word limits, figure count, reference style, reporting-guideline requirements, data-availability policy, preprint policy) and canonical author-instructions URLs for Nature Commun, Leukemia, Genome Medicine, Cell Reports Medicine, Blood Cancer Journal, Briefings in Bioinformatics.

Copy the LaTeX skeleton from assets/latex/ and edit the bibliography style, section order, and figure-count cap to match the target journal.

Exit criterion: journal named, author-instructions distilled into manuscript/JOURNAL.md, LaTeX template adjusted, tex compiles with placeholder title.

Phase 4 - Project setup + preregistration + method design

Scaffold with scripts/init_project.py <project-dir> (creates data/raw, data/processed, data/results, analysis/, manuscript/, figures/, docs/, .github/workflows/, .gitignore, LICENSE, README.md, pyproject.toml, and a docs/prereg.md stub). The script refuses to overwrite existing critical files without --force (which takes timestamped backups). Install Python dependencies with uv add.

Commit docs/prereg.md before touching any outcome-related analysis. The preregistration commit is the single most important integrity artefact; skipping it turns any later outcome pivot into HARKing. See references/preregistration-and-integrity.md for the required contents and the rules that make the orthogonal-outcome pivot legitimate.

Design the method before looking at results. See references/method-design.md for the rigor checklist (leakage-free feature selection, stability / parameter sweep, permutation null, proportional-hazards tests, nested C-index, calibration + DCA / NRI if clinical, bootstrap CIs, multiple-testing correction). The target journal's author instructions drive the checklist priority - clinical journals weight calibration + DCA; methods journals weight stability + null models.

Common open-dataset URLs and download recipes live in references/open-datasets.md; access codes (Open / Gated / Mixed) in references/data-first-novelty.md tell you which raw artefacts can be redistributed in the final release.

Exit criterion: docs/prereg.md committed, data downloaded, analysis scripts stubbed, uv run python analysis/01_prepare_data.py succeeds.

Phase 5 - Analysis, figures, manuscript with journal-matched writing

Iterate analysis -> figures -> LaTeX. Structural conventions in references/manuscript-structure.md.

Writing style is journal-specific. See references/reader-reviewer-writing.md for the explicit reader and reviewer profiles of Nature Commun, Leukemia, Genome Medicine, BCJ, Cell Reports Medicine, Briefings in Bioinformatics, and the abstract-headline + discussion-caveat patterns that work at each. A Leukemia reader expects the clinical hook in the first sentence; a Briefings reader expects the methods gap; a Nature Commun reader expects the broad-interest statement followed by the quantitative headline.

Compile with cd manuscript && make.

Exit criterion: full PDF compiles; abstract-headline, discussion-caveats, and data/code availability sections match the style targeted at manuscript/JOURNAL.md.

Phase 6 - Reviewer cycle

Dispatch four adversarial reviewers in parallel (methods, clinical, biostatistics, target-journal editor). Iterate until unanimous accept. Pattern and prompt templates in references/reviewer-cycle.md. The target-journal editor reviewer is instantiated with that journal's historical concerns - loaded from the author-instructions file produced in Phase 3.

Phase 7 - Repo and tagged release

Create private GitHub repo, wire CI, push tag, release PDF + .tex + source bundle. Exact commands + CI workflow template in references/release-workflow.md. CI template at assets/github/release.yml; LaTeX helpers at assets/latex/Makefile + latexmkrc. Use scripts/new_release.sh for a preflight-checked release (rejects dirty git tree, missing gh auth, stale placeholder DOIs, absent figures).

Release naming: v<MAJOR>.<MINOR>.<PATCH> where MAJOR bumps reflect scientific pivots, MINOR bumps reflect reviewer-round revisions, PATCH bumps reflect typos and terminology.

Phase 8 - Submission support (optional but expected at IF 10+)

The skill's release workflow ends at a tagged preprint-ready PDF. Actual submission happens off-skill, but the supporting artefacts are templated here:

• Preprint posting (bioRxiv / medRxiv / arXiv) + Zenodo DOI minting: references/preprint-and-dois.md. Ship the preprint before the journal submission so the cover letter can cite it.
• Cover letter + point-by-point rebuttal templates: references/cover-letter-and-rebuttal.md. Rebuttals are the highest-leverage artefact in a revision cycle.
• Submission-portal quirks (Editorial Manager, ScholarOne, Snapsubmit, figure-format, ORCID, supplementary material): references/submission-portals.md.
• CRediT authorship, ethics and competing-interests language: references/credit-and-ethics.md. Do the CRediT assignment with all authors during Phase 4 or 5; negotiating at the last minute causes submission delays.

Scripts

• scripts/init_project.py <project-dir> - scaffold directory tree, copy LaTeX + GH Actions templates, write .gitignore, LICENSE, README.md, pyproject.toml. Run once at start of Phase 4.
• scripts/new_release.sh <project-dir> <version> "<title>" "<notes>" - build PDF, stage artefacts into /tmp, create tagged gh release with PDF + .tex + source bundle. Run at Phase 7 and after every review round.

Non-goals

• Wet-lab experiments, patient-level trial design, regulatory submission.
• Actual journal submission (the skill ends at a release-tagged preprint-ready PDF matched to one target journal's style).
• Author-name or ORCID fill-in (always prompt the user).

Pitfalls the skill's references protect against

• Topic-first workflow that picks a fashionable method and shops for data -> Phase 1 is data-first precisely to avoid this.
• Novelty hallucinated from training data -> novelty-search.md verifies against current literature.
• Picking the highest-IF journal before checking acceptance probability -> journal-targets.md scores fit, not just IF.
• Writing manuscript before reading author instructions -> Phase 3 fetches and distils them up front.
• Gene-panel leakage across cohorts -> method-design.md.
• Claims that collapse under a proper clinical baseline -> reviewer-cycle.md catches it before submission.
• Release without GPG-signed tag fails on first push -> release-workflow.md uses gh release create (remote tag) to sidestep local signing.