write-report-zeroshot

write-report-zeroshot

Two-step skill: (1) call MI2 to get per-finding probabilities for the image, then (2) hand those off to the existing write-report-simple skill to produce the actual report. The report itself never mentions MI2 — it reads as the radiologist's direct read of the image.

Budget

• Tool calls: 1 — exactly one mcp_medimageinsig_zeroshot_classify call per image.
• Output: ≤ 200 words for the full report.
• Do not re-invoke skills already used in this turn.
• Do not retry tool calls on success.

How to use

1. Resolve the image to an absolute path. The MCP tool runs in a different working directory than the agent. If image_path is not absolute, resolve it (e.g. via realpath) before calling the tool. If you can't resolve it, ask the user.

2. Get scores from MI2.

   • Read the canonical label list from references/labels.md and use those strings verbatim.
   • If the caller provided extra_labels, append them.
   • Call mcp_medimageinsig_zeroshot_classify with file_path = the absolute path from step 1 and labels = the assembled list.
   • Bucket the returned scores using the cutoffs below.

3. Hand off to write-report-simple. Invoke the write-report-simple skill (already in the workspace skill drawer at .github/skills/write-report-simple/). Pass:

   • findings: the bucketed list (positives + possibles), using the original MI2 label names.
   • scores: the raw per-finding probabilities (so the report skill can calibrate hedging).

   The downstream skill derives confidence from scores.

4. Return what write-report-simple returns. Do not re-wrap, re-format, or annotate the report. The downstream skill emits the <confidence> and <report> blocks; pass them through unchanged.

Score buckets

Probabilities sum to ~1.0 across the label set. With the default label list, use these cutoffs:

• ≥ 0.20 → positive finding (pass to write-report-simple as a confident finding).
• 0.08 – 0.20 → possible finding (pass through; the report skill hedges accordingly).
• < 0.08 → absent (omit from findings).

Critical rule

The report must not mention MI2, zero-shot scores, or any model output. That's the responsibility of write-report-simple (which has the rule baked in), but it's worth restating here: this skill exists to ground the report in vision-model evidence, not to surface that evidence in the prose. If you want to attribute or caveat the model in the chat reply outside the <report> tags, that's fine.