Skills & Report Quality
Claude skills and LLM-backed tools for scoring, checking, and rewriting pathology reports
https://github.com/sbalci/pathology-report-checker-skill
https://github.com/sbalci/pathology-report-colon-QA
https://github.com/sbalci/pathology-report-checker-skill
Skill catalog
| Skill | Purpose | Repo |
|---|---|---|
| Report checker | CAP/ICCR compliance scoring of synoptic elements | pathology-report-checker-skill |
| Colon QA | Narrative-to-synoptic gap analysis for colon resections | pathology-report-colon-QA |
| Pathology skills collection | Summaries, translations, lay-language rewrites | pathology-skills-collection |
| Scientific skills | Literature search, citation, figure generation | claude-scientific-skills |
| n8n pathology | Workflow glue across the skills above | n8n-pathology |
When to use which
- Compliance scoring on a batch of reports →
pathology-report-checker-skill. - Investigating a specific organ’s reporting quality → organ-specific QA skill (colon is the reference; others in flight).
- Writing lay summaries or translating for patients →
pathology-skills-collection. - Literature / citation work embedded in a manuscript →
claude-scientific-skills. - Chaining several skills together on a batch →
n8n-pathologyworkflows.
Evaluation
Every skill ships with:
- A fixture set of redacted sample reports.
- A scoring rubric with ground-truth labels agreed by two pathologists.
- A CI run that fails the build if rubric accuracy drops below the documented floor.
How to contribute
- Pick a skill and open its
README. - Run the fixture eval; read failure cases.
- Propose one new test case from a real (redacted) failure mode you’ve seen.
- Open a PR with the fixture and, if needed, a prompt / rule update.
Pitfalls
- Do not send raw (un-redacted) reports through any LLM skill. The staging pipeline in
extract-report-textis where redaction happens. - Skills are tuned to specific Claude model versions. When a new model is released, re-run the full rubric before rolling it out.