Review Viewer

Yao Meta Skill

Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.

maturity: governed archetype: governed format: agent-skills updated: 2026-03-31 intent confidence: 100 / 100

Architecture at a glance

Inputs

workflow, prompt, transcript, docs, or notes

Boundary

Logic

For one-off/no reusable process: `Do not create a skill`; `near-neighbor`; require `repeated use` + `reusable output contract`.; Capture job, output, exclusions, constraints, standards, and the lightest fit.; Scan references in order: external benchmark, user source, local fit; surface only uncertainty or conflict.

Usage

Use $yao-meta-skill to turn my workflow or notes into a reusable skill with lean structure, clear triggering, and the right evals.; Use this skill when the request matches: Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.

Review the top iteration directions before growing the package.

Core logic

For one-off/no reusable process: `Do not create a skill`; `near-neighbor`; require `repeated use` + `reusable output contract`.
Capture job, output, exclusions, constraints, standards, and the lightest fit.
Scan references in order: external benchmark, user source, local fit; surface only uncertainty or conflict.
Write `description` early, test route quality, then add only earned folders and gates.
Add output-risk, artifact-design, prompt-quality, system-model, and next directions only when useful.

How to use it

Use $yao-meta-skill to turn my workflow or notes into a reusable skill with lean structure, clear triggering, and the right evals.
Use this skill when the request matches: Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes. Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.

Intent questions

If this skill worked beautifully, what recurring job would it reliably handle for the user every time?
This reveals the real job-to-be-done and gives the package a humane center instead of a guessed prompt shape.
When someone reaches for this skill in the real world, what materials will they actually hand to it?
Input shape decides whether references, scripts, or templates are needed.
What finished output should it hand back so the user can immediately keep moving?
Outputs should drive the package structure before extra guidance is added.
Which nearby requests should this skill politely refuse so the boundary stays clean?
The exclusion list is the fastest route to better trigger quality.
What matters most here: speed, consistency, auditability, portability, governance, or tone/style fit?
Constraints decide how much structure, packaging, and review the skill actually needs.

Why this package is strong

触发面保持精简，并锚定在 frontmatter description。
已生成 Skill IR，核心语义可先于平台打包被审查和迁移。
已生成目标编译报告，可审查 IR 到 OpenAI、Claude、generic 等目标契约的映射。
已生成 Output Eval Lab scorecard，可比较 with-skill 与 baseline 输出质量。
已生成 Output Execution Runs，可区分记录样本、命令执行和模型执行证据。
已生成 Output Review Adjudication，可记录盲评决策、一致率和待评审项。

Borrow plan

No external benchmark objects recorded yet. Add 2 to 5 references before deepening the package.

Compare view

Winner: Current

{'priority': ['fewest false positives', 'fewest false negatives', 'highest near-neighbor pass rate', 'highest negative pass rate', 'highest precision', 'highest recall', 'shortest description']}

Variant	Tokens	Strategy
Baseline	8	existing
Current	53	current
Current	53	current

Variant diff studio

Baseline

existing

Create and improve agent skills.

tokens 8 (0)dev 0 (0)holdout 0 (0)

Adds relative to baseline

Create and improve agent skills.

Drops from baseline

No dropped cues.

Current

current

tokens 53 (+45)dev 0 (0)holdout 0 (0)

Adds relative to baseline

Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes.
Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.

Drops from baseline

Create and improve agent skills.

Winner — Current

current

tokens 53 (+45)dev 0 (0)holdout 0 (0)

Adds relative to baseline

Create, refactor, evaluate, and package agent skills from workflows, prompts, transcripts, docs, or notes.
Use for skill creation, reusable workflow packaging, skill improvement, evals, and team-ready distribution.

Drops from baseline

Create and improve agent skills.

Evidence readiness

Readiness score: 85/100

Intent clarity · ready
100/100 intent confidence.
Benchmark coverage · ready
3 GitHub benchmark repositories attached.
Pattern gate · ready
5 accepted, 1 deferred using threshold 4/4.
Conflict handling · decision needed
The stated preference leans lightweight or speed-first, while the benchmark mix leans toward governance, review, or heavier evaluation structure.
Output risk profile · ready
5 output risk families attached.
Artifact design profile · ready
Metric-first dashboard with stable dimensions, short labels, visible deltas, and narrative callouts only where they change interpretation.
Prompt quality profile · ready
89.0/100 prompt-facing quality score.

Honest boundary check

Are the known limits visible before the package deepens?
Does the evidence support the borrowed patterns?
Should uncertainty become a clarification question instead of more structure?

Output risk profile

Markdown readability
Tables can render as dense grids with weak hierarchy or poor mobile readability.; Long bullets can make the output look complete while hiding the actual decision logic.
Citation and footnote clutter
Footnote markers or dense citation notes can interrupt the reading flow.; Evidence can be over-attached to obvious statements and under-attached to risky claims.
Screenshot and visual capture
Screenshots can be captured from the wrong state, wrong viewport, or wrong crop.; Missing screenshots can cause the skill to invent visual references instead of declaring the gap.

Self-repair checks

Preview whether each table still reads well when columns are narrow.
Convert any table with paragraph-length cells into bullets or cards.
Remove decorative citations that do not support a material claim.
Move repeated source explanations into one compact source note.
Check that every screenshot reference points to a real provided or generated asset.

Artifact design profile

Design system: metric editorial

Dashboard or metrics page
Metric-first dashboard with stable dimensions, short labels, visible deltas, and narrative callouts only where they change interpretation.
Review viewer
Side-by-side reviewer studio with explicit tradeoffs, evidence readiness, and fast paths for approving, blocking, or requesting one focused fix.
Code, CLI, or implementation guide
Execution-focused technical artifact with environment assumptions, copyable commands, expected outputs, and side effects made explicit.

Visual quality gates

Avoid paragraph-heavy table cells.
Keep charts tied to one analytical question each.
Preserve stable color meaning across metrics and entities.
Make differences visible instead of hiding them in prose.
Separate author-facing recommendations from reviewer-only evidence.

Prompt quality profile

Relevance: prompt-aware · score 89.0 / 100 · complexity expert

Completeness · 100 / 100
Name missing inputs, outputs, constraints, or success standards before deepening the package.
Clarity · 85 / 100
Replace broad verbs with observable actions and define what done means.
Consistency · 90 / 100
Check that role, task, format, exclusions, and examples do not contradict each other.
Practicality · 90 / 100
Add runnable steps, examples, or verification cues instead of abstract advice.
Specificity · 80 / 100
Anchor wording in the user's audience, domain nouns, and target outcome.

RTF to skill mapping

Role
Use an operator role with explicit boundaries, inputs, outputs, and failure handling.
Task
Convert the job into ordered steps with validation checks and stop conditions.
Format
Return a runbook-like handoff with commands, checks, owners, and next actions when relevant.

Reference coach

obra/superpowers

Borrow now

Borrow the way it turns a messy workflow into a repeatable operating path.
Borrow the clear execution entrypoints and command structure.

Avoid

Do not import process overhead that only exists for that project's scale.

affaan-m/ECC

Borrow now

Borrow the way it turns a messy workflow into a repeatable operating path.
Borrow the clear execution entrypoints and command structure.

Avoid

Do not import process overhead that only exists for that project's scale.

multica-ai/andrej-karpathy-skills

Borrow now

Borrow explicit validation and quality gates that make iteration safer.
Borrow the way it separates explanation, examples, and reusable structure.

Avoid

Do not clone heavyweight evaluation scaffolding if a lighter gate is enough here.

Decide before you deepen

Choose one pattern to borrow on purpose, not three at once.
State one thing this skill will not inherit from the benchmark objects.
Only deepen the package after that choice is visible in the boundary or execution flow.

Reference synthesis

Official skill anatomy and context discipline

Borrow now

Borrow progressive disclosure: keep the entrypoint lean and move depth into references or scripts.

Avoid

Do not let packaging or platform concerns swallow the core job boundary.

Human-in-the-loop verification

Borrow now

Borrow a review checkpoint wherever trust matters more than raw speed.

Avoid

Do not force every skill through heavyweight review when the risk is low.

Boundary-first design

Borrow now

Borrow the discipline of defining what the skill should not own before growing the package.

Avoid

Do not expand execution assets until route boundaries stay clean.

Borrow now

Borrow progressive disclosure: keep the entrypoint lean and move depth into references or scripts.
Borrow a review checkpoint wherever trust matters more than raw speed.
Borrow the discipline of defining what the skill should not own before growing the package.
Borrow the way it turns a messy workflow into a repeatable operating path.

Use the recommendation by default. Only surface the underlying benchmark tradeoffs when intent is uncertain or a real design conflict needs a deliberate call.

Top three next moves

Borrow one proven pattern on purpose

You already have public benchmark objects. The next gain is to choose one pattern intentionally instead of absorbing everything loosely.

Read the strongest pattern from obra/superpowers.
Decide whether to borrow method, structure, execution, or portability, but only one of them first.
Record what you will not borrow so the package stays light.

Unlocks: A cleaner package shape with less accidental over-design.

Harden portability semantics

The skill already signals reuse across environments, so contract clarity matters early.

Confirm activation mode, execution context, and trust assumptions.
Add or review degradation strategy for non-native targets.
Package the skill once to verify adapter expectations.

Unlocks: Safer cross-environment reuse with less target drift.

Create an iteration evidence loop

The package should show what changed and why after the first draft.

Generate the HTML skill report and keep it aligned with the package.
Record reference scan choices and non-goals.
Capture the next iteration choice explicitly before adding more files.

Unlocks: A clearer path for the next author or reviewer.

Recent feedback

No lightweight feedback captured yet. Use `yao.py feedback` to record quick review notes.

Promotion status

Promote: 0

Keep current: 3

Blocked: 0

Package map

SKILL.md — Skill entrypoint
README.md — Human-readable usage guide
agents/interface.yaml — Neutral interface metadata
manifest.json — Lifecycle and portability metadata
references — Extended guidance and reusable notes
scripts — Deterministic helpers or local tooling
evals — Trigger and quality checks
reports — Generated evidence and overview artifacts

First-pass review frame

Does the trigger stay narrow enough for the intended job?
Does the archetype match the real reuse level?
Are we adding structure faster than we are adding reliability?
Should the next step be trigger tightening, execution assets, or portability hardening?

Authoring discipline

Name unresolved assumptions before deepening the package.
Keep the package no larger than the recurring job requires.
Touch only files that directly support the requested change.
Tie every meaningful new artifact to a check or reviewer note.

Reviewer guardrails

Block speculative features that are not backed by real workflow variation.
Move unverifiable ideas into next-step candidates instead of baseline structure.
Reject decorative folders, reports, or governance that do not reduce risk.
Ask for one high-leverage clarification when job, output, or exclusion is still fuzzy.