feat: harden SKILL.md evidence structure into an explicit output format - #17
Merged
Merged
Conversation
Both the Haiku 4.5 and Sonnet 5 rubric runs found a consistent instruction-following gap on scenario A (Stack Trace Anchoring) and, more mildly, scenario B: with-skill responses often skipped the skill's own FACT/ASSUMPTION/UNKNOWN labeling and confidence classification when given a bare stack trace or crash report, while unprimed baseline responses to the identical prompt frequently produced that structure spontaneously. Both prior reports recommended converting the prose guidance into an explicit, harder-to-skip output format. Rewrites "Establish What Is Known" and "Classify the Conclusion" to require an exact FACTS/ASSUMPTIONS/INFERENCES/UNKNOWNS/CONFLICTING EVIDENCE/CLAIMS block and an explicit "Conclusion: VERIFIED | STRONGLY SUPPORTED | INFERRED" line for any non-trivial diagnosis, and states explicitly that a bare stack trace or crash report is non-trivial by default even when the eventual fix is small. Adds a matching "Never" entry. No other section changed. Verified against Sonnet 5: scenario A went from 8.0/20 to 19/20 and scenario B from 11.0/20 to 20/20 on the same rubric (both now above their baseline scores of 16.5 and 16.0). Spot-checked scenario E (trivial edit) to confirm the tightened format doesn't add unwanted ceremony to genuinely trivial tasks — it doesn't.
4 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
FACTS:/ASSUMPTIONS:/INFERENCES:/UNKNOWNS:/CONFLICTING EVIDENCE:/CLAIMS:block and a requiredConclusion: VERIFIED | STRONGLY SUPPORTED | INFERREDline for any non-trivial diagnosis, with an explicit statement that a bare stack trace/crash report counts as non-trivial by default.SKILL.mdchanged.Test plan
tests/results/2026-08-21-skillmd-format-hardening.md, not done in this PR (this PR verifies the specific fix, not a full re-characterization).