omics-skills AGENTS.md

AGENTS.md

Guidance for AI coding agents (Claude Code, Codex CLI, Cursor, Copilot) working in this repository.

Related docs: docs/CONTRIBUTING.md (contributor workflow), docs/development.md (validation commands), docs/INSTALL.md (installation).

---

Repository overview

4 agents and 34 skills covering bioinformatics, literature discovery, scientific writing, and data visualization. Runs under Claude Code and the Codex CLI.

Layout:

• agents/ — 4 agent definitions (markdown)
• skills/ — skill directories; each has a SKILL.md
• scripts/ — router, catalog builder, hook, installer, benchmark
• tests/ — unit tests + routing benchmark
• catalog/ — generated routing artifact (catalog.json)
• docs/ — MkDocs site sources, routing model, benchmark baseline
• Makefile — install, catalog, hook, benchmark, uninstall targets

Install: make install symlinks agents and skills into ~/.claude/ and ~/.codex/.

---

Default Workflow Selection

When working in this repository without a specialized agent prompt, do not choose skills ad hoc.

Start with the catalog:

python3 scripts/skill_index.py route "<task>"

Use the returned agent, primary skills, and suggested order as the default workflow. Then open the referenced agent file in agents/ and the referenced SKILL.md files before proceeding. Only deviate from the returned path when the request clearly falls outside the suggested workflow.

For installed environments outside the repository checkout, use:

python3 ~/.agents/omics-skills/skill_index.py route "<task>"

---

Scientific Workflow Guardrails

For omics or scientific project work, agents must maintain an explicit reasoning loop instead of treating analysis steps as a linear pipeline.

1. Hypothesis register: before the first analysis step, create at least 5 distinct working hypotheses or explanations. Include technical artifacts and null explanations alongside biological mechanisms. If fewer than 5 are plausible, state why and add discriminating negative controls or failure modes.
2. Intermediate reflection: after each major intermediate result or QC gate, write a short reflection covering what was observed, whether the result passed QC, which hypotheses gained or lost support, what alternative explanations remain, and the next discriminating check.
3. Literature context: after the initial hypothesis register and after any unexpected, central, or final finding, run an additional literature search using polars-dovmed or another appropriate literature-search skill. Use broad synonym-aware queries, summarize the relevant evidence with DOI/PMCID when available, and state whether the literature supports, contradicts, or narrows each leading hypothesis.
4. Hypothesis revision: revise and rank the hypothesis register as evidence accumulates. Do not silently discard hypotheses; mark them as supported, weakened, ruled out, or unresolved with the evidence that changed the status. Keep at least 5 active hypotheses while the project is exploratory; generate replacements when hypotheses are ruled out.
5. Final synthesis: final reports must include hypotheses considered, intermediate reflections, literature context, revised hypothesis ranking, and the next experiments or analyses that would best separate remaining alternatives.