yaojingang
4.1 KiB
Output Blind A/B Review Pack
This packet hides whether each variant came from the baseline or the skill-guided output. Use the separate answer key only after review.
- Pairs:
5 - Seed:
yao-output-eval-blind-v1 - Answer key separate:
True
Case: skill-package-contract
Prompt: Turn this repeated workflow into a reusable team skill package.
Rubric:
has-entrypoint(1.0): Output names the routeable Skill entrypoint.has-interface(1.0): Output includes neutral interface metadata.has-report(1.0): Output points reviewer to generated overview report.has-resource-boundary(1.0): Output preserves entrypoint/resource separation.
Variant A
Create a routeable SKILL.md, agents/interface.yaml, reports/skill-overview.html, and a focused eval plan. Keep the root entrypoint lean, move durable guidance into references/, scripts into scripts/, and evidence into reports/.
Variant B
I can write a prompt for that workflow and include a short checklist.
Case: output-eval-expectation
Prompt: Upgrade this production skill so we know whether the generated output is better than baseline.
Rubric:
has-baseline-comparison(1.0): Output explicitly compares with-skill and baseline outputs.has-assertion-grading(1.0): Output includes objective assertion grading.has-scorecard(1.0): Output produces a scorecard report path.has-failure-taxonomy(1.0): Output records failure taxonomy.
Variant A
Add Output Eval Lab cases with baseline_output and with_skill_output, run assertion grading, report with-skill vs baseline pass-rate delta, and record failure taxonomy plus next fixes in reports/output_quality_scorecard.md.
Variant B
Add more examples and run the trigger eval again.
Case: ir-before-packaging
Prompt: Package this skill for OpenAI, Claude, Agent Skills, and generic targets.
Rubric:
has-ir-first(1.0): Output requires Skill IR before packaging.has-core-ir-fields(1.0): Output lists core IR contract fields.has-targets(1.0): Output names requested runtime targets.has-semantic-preservation(1.0): Output says targets must preserve the capability contract.
Variant A
Export folders for each platform and adjust files as needed.
Variant B
Export Skill IR first with job_to_be_done, trigger_surface, workflow, resources, eval_plan, risk, and governance. Then compile or package targets from the IR so OpenAI, Claude, Agent Skills, and generic packages preserve the same capability contract.
Case: near-neighbor-boundary
Prompt: I only need a one-off summary of these notes, not a reusable process.
Rubric:
declines-skill(1.0): Output refuses unnecessary skill creation for one-off work.names-near-neighbor(1.0): Output labels the request as near-neighbor instead of owned work.requires-reuse-signal(1.0): Output asks for repeat-use evidence before packaging.
Variant A
Create a SKILL.md and a reusable workflow anyway so future notes can use it.
Variant B
Do not create a skill for this one-off request. Treat it as a near-neighbor: answer the summary directly unless the user confirms repeated use, shared ownership, or a reusable output contract.
Case: file-backed-governed-package
Prompt: Turn the attached release brief source into a governed skill package.
Rubric:
uses-file-backed-evidence(1.0): Output names file-backed source evidence.has-governance(1.0): Output preserves governed ownership metadata.has-output-contract(1.0): Output preserves output and rollback boundaries.has-trust-and-scorecard(1.0): Output requires trust and output scorecard artifacts.does-not-invent-evidence(1.0): Output forbids invented launch evidence.
Variant A
Draft a release announcement with the changelog, support notes, and owner name.
Variant B
Use the file-backed fixture as source evidence, then create a governed skill package with SKILL.md, agents/interface.yaml, owner, review cadence, input_files, output contract, rollback boundary, trust report, and reports/output_quality_scorecard.md. Mark missing launch metrics as missing evidence instead of inventing them.