| name | schema-v1-smoke-human-review |
| description | Human review workflow for schema-v1 smoke outputs before go/no-go decisions. |
| trigger | Use when a reviewer validates schema-v1 smoke outputs for parse status, release threshold, model identity, and quant labels. |
| auto_generated | {"by":"skill-opt","date":"2026-05-31T00:00:00.000Z","source_goal":"Generate a human review workflow for schema-v1 smoke outputs, covering parse_ok, go/no-go threshold, model identity, and quant labeling.","planner_model":"codex-cli/default","review_required":true} |
schema-v1-smoke-human-review
When to use
Use this skill when reviewing schema-v1 smoke output artifacts before accepting, publishing, or using them as gate evidence. It applies when a human must confirm that the output parsed successfully, meets the go/no-go threshold, identifies the model correctly, and labels quantitative results consistently.
Steps
- Locate the schema-v1 smoke output artifact and confirm it is the exact artifact under review, not a stale or regenerated file.
- Verify the artifact declares schema-v1 and contains the expected top-level fields for the smoke output format.
- Check
parse_ok first:
- If
parse_ok is not present, not boolean, or false, mark the review as no-go.
- Record the parse failure reason if the artifact provides one.
- Review the go/no-go threshold:
- Identify the configured threshold used for the smoke run.
- Compare the observed score or pass count against that threshold.
- Mark go only when the threshold is explicitly met or exceeded.
- Mark no-go when the threshold is missing, ambiguous, or failed.
- Validate model identity:
- Confirm the model name/id is present.
- Confirm it matches the expected model for the run.
- Flag aliases, missing versions, or unexpected providers for follow-up before go.
- Validate quant labeling:
- Confirm quantitative fields are labeled with clear metric names.
- Confirm units, percentages, counts, or score ranges are explicit.
- Confirm labels distinguish raw values from normalized, averaged, or thresholded values.
- Review any warnings, skipped checks, or partial results and decide whether they affect the gate.
- Record the final human decision as
go or no-go with a short reason tied to parse_ok, threshold result, model identity, and quant labeling.
- If no-go, provide the minimum corrective action needed before re-running or re-reviewing the smoke output.
DoD
Anti-patterns
- Treating a successful file read as equivalent to
parse_ok: true.
- Approving a run when the threshold is missing or inferred from memory.
- Ignoring model aliases, missing versions, or unexpected provider names.
- Accepting unlabeled numbers without units, score ranges, or metric names.
- Mixing raw counts, percentages, normalized scores, and threshold decisions under one label.
- Re-reviewing a regenerated artifact while claiming it is the original smoke output.