YAO
31ce04c655
Merge the beta-ready Yao Meta Skill architecture, report, evidence gate, and release-boundary updates.\n\nRelease boundary: beta/public testing is allowed; formal world-class, fully reviewed, or superiority claims remain blocked until the pending evidence gates are accepted.
181 lines
9.4 KiB
JSON
181 lines
9.4 KiB
JSON
{
|
|
"skill_name": "yao-meta-skill",
|
|
"description": "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.",
|
|
"systems_doctrine": "Structure drives behavior: improve the boundary, feedback loops, drift watch, and leverage points before adding weight.",
|
|
"stability": {
|
|
"score": 100,
|
|
"band": "system-ready"
|
|
},
|
|
"boundary_map": {
|
|
"owned_job": "Turn repeated workflows, prompts, transcripts, runbooks, documents, or existing skill packages into routeable, evaluable, packageable, and governable agent skills for personal, team, library, or governed reuse.",
|
|
"input_boundary": [
|
|
"rough workflow notes, SOPs, runbooks, prompts, transcripts, documents, or repeated task descriptions",
|
|
"an existing skill directory that needs refactor, evaluation, packaging, or governance hardening",
|
|
"target platform requirements such as OpenAI, Claude, generic Agent Skills, or team distribution",
|
|
"benchmark references, local constraints, desired maturity tier, and review standards"
|
|
],
|
|
"output_boundary": "A working skill package with lean SKILL.md, aligned agents/interface.yaml, justified references, scripts only when useful, eval evidence, reports, packaging metadata, and clear next iteration recommendations.",
|
|
"non_goals": [
|
|
"one-off writing, translation, explanation, or brainstorming requests that do not need a reusable skill",
|
|
"general code review or debugging unless the user is packaging that workflow as a skill",
|
|
"raw private material that was not intentionally supplied as skill evidence",
|
|
"platform-specific plugin creation when the request is not about an agent skill package"
|
|
],
|
|
"constraints": [
|
|
"keep SKILL.md lean and route primarily through frontmatter description",
|
|
"put durable guidance in references, executable logic in scripts, and evidence in reports",
|
|
"default to the lightest reliable mode before adding governance weight",
|
|
"preserve portability across OpenAI, Claude, generic, and Agent Skills compatible targets",
|
|
"avoid raw prompt, output, transcript, or private content in telemetry"
|
|
],
|
|
"standards": [
|
|
"trigger boundaries must be tested with should-trigger and should-not-trigger cases",
|
|
"production and higher maturity work needs output eval, trust, runtime conformance, and Review Studio evidence",
|
|
"governed work needs owner, review cadence, permission approvals, registry metadata, package verification, and install simulation",
|
|
"generated reports should be bilingual or reviewer-friendly when they are user-facing",
|
|
"each new asset must earn its place by reducing ambiguity, risk, or repeated work"
|
|
],
|
|
"human_judgment_boundary": [
|
|
"Ask one focused clarification when the real job, output, or exclusion boundary is unclear.",
|
|
"Escalate visible tradeoffs when benchmark patterns conflict with local privacy, naming, or governance constraints.",
|
|
"Do not silently broaden the skill into adjacent jobs just because the examples are nearby."
|
|
],
|
|
"maturity_assumption": "governed"
|
|
},
|
|
"feedback_loops": [
|
|
{
|
|
"name": "Intent boundary loop",
|
|
"signal": "Intent confidence score is 100/100.",
|
|
"response": "Ask only the highest-leverage clarification before adding package weight.",
|
|
"evidence": "reports/intent-confidence.md and reports/intent-dialogue.md"
|
|
},
|
|
{
|
|
"name": "Reference synthesis loop",
|
|
"signal": "Benchmark patterns are useful only after they are abstracted into borrow and avoid guidance.",
|
|
"response": "Borrow one pattern at a time and keep the rest as reviewer-visible evidence.",
|
|
"evidence": "reports/reference-synthesis.md",
|
|
"current_patterns": [
|
|
"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.",
|
|
"Borrow the clear execution entrypoints and command structure."
|
|
]
|
|
},
|
|
{
|
|
"name": "Output quality loop",
|
|
"signal": "Generated output may fail in recurring domain-specific ways.",
|
|
"response": "Apply predicted output-risk families as self-repair checks before final output.",
|
|
"evidence": "reports/output-risk-profile.md",
|
|
"current_risk_families": [
|
|
"Markdown readability",
|
|
"Citation and footnote clutter",
|
|
"Screenshot and visual capture",
|
|
"Code and command safety",
|
|
"Tone and specificity"
|
|
]
|
|
},
|
|
{
|
|
"name": "Reviewer feedback loop",
|
|
"signal": "Human review catches drift that static checks miss.",
|
|
"response": "Capture lightweight feedback and turn repeated findings into gates or references.",
|
|
"evidence": "reports/review-viewer.html and feedback records"
|
|
},
|
|
{
|
|
"name": "Lifecycle loop",
|
|
"signal": "As reuse grows, the skill needs stronger gates, ownership, and regression evidence.",
|
|
"response": "Promote only when the next gate improves reliability more than context cost.",
|
|
"evidence": "manifest.json, reports/iteration-directions.md, and governance checks"
|
|
}
|
|
],
|
|
"drift_watch": [
|
|
{
|
|
"name": "Trigger drift",
|
|
"watch_signal": "Users start invoking the skill for adjacent one-off or explanation-only requests.",
|
|
"countermeasure": "Add near-neighbor exclusions and route evals before expanding workflow steps.",
|
|
"cadence": "per trigger or description change"
|
|
},
|
|
{
|
|
"name": "Output drift",
|
|
"watch_signal": "Outputs remain valid but become generic, cluttered, or weakly aligned with the user's domain.",
|
|
"countermeasure": "Refresh output-risk and artifact-design profiles, then add one self-repair check.",
|
|
"cadence": "after the first 3-5 real uses",
|
|
"risk_families": [
|
|
"Markdown readability",
|
|
"Citation and footnote clutter",
|
|
"Screenshot and visual capture",
|
|
"Code and command safety",
|
|
"Tone and specificity"
|
|
]
|
|
},
|
|
{
|
|
"name": "Reference drift",
|
|
"watch_signal": "Borrowed benchmark patterns no longer fit the local job or add ceremony without payoff.",
|
|
"countermeasure": "Re-run reference synthesis and keep only patterns that improve the current boundary.",
|
|
"cadence": "per material benchmark or product assumption change"
|
|
},
|
|
{
|
|
"name": "Governance drift",
|
|
"watch_signal": "Skill usage becomes team-critical while ownership, review cadence, or rollback evidence stays informal.",
|
|
"countermeasure": "Promote maturity tier and add reviewer-visible lifecycle evidence.",
|
|
"cadence": "monthly"
|
|
}
|
|
],
|
|
"failure_pattern_map": [
|
|
{
|
|
"family": "Boundary failure",
|
|
"symptom": "The skill handles nearby requests that were never part of the recurring job.",
|
|
"repair": "Narrow the description and add explicit non-goals before adding more execution steps."
|
|
},
|
|
{
|
|
"family": "Feedback gap",
|
|
"symptom": "The skill has rules but no signal telling authors which rule should change after use.",
|
|
"repair": "Turn repeated reviewer feedback into one eval, one reference note, or one self-repair check."
|
|
},
|
|
{
|
|
"family": "Output degradation",
|
|
"symptom": "The result is structurally correct but generic, cluttered, or weakly matched to the user's domain.",
|
|
"repair": "Use output-risk families as pre-final checks.",
|
|
"current_risk_families": [
|
|
"Markdown readability",
|
|
"Citation and footnote clutter",
|
|
"Screenshot and visual capture",
|
|
"Code and command safety",
|
|
"Tone and specificity"
|
|
]
|
|
},
|
|
{
|
|
"family": "Prompt-behavior mismatch",
|
|
"symptom": "The role, task, and format are copied from a prompt instead of becoming stable skill behavior.",
|
|
"repair": "Convert reusable role/task/format assumptions into workflow, reports, or references.",
|
|
"watch_axes": []
|
|
}
|
|
],
|
|
"leverage_points": [
|
|
{
|
|
"rank": 2,
|
|
"point": "Tune the frontmatter description",
|
|
"why": "The description is the highest-leverage routing surface.",
|
|
"move": "Name the recurring job, expected input, output, and strongest non-goal in compact language."
|
|
},
|
|
{
|
|
"rank": 3,
|
|
"point": "Install output self-repair checks",
|
|
"why": "The likely failure families are: Markdown readability, Citation and footnote clutter, Screenshot and visual capture.",
|
|
"move": "Add only the checks that prevent recurring output mistakes."
|
|
},
|
|
{
|
|
"rank": 4,
|
|
"point": "Borrow one pattern, not a whole product",
|
|
"why": "External references improve quality when reduced to structure, not copied as surface style.",
|
|
"move": "Start from: Borrow progressive disclosure: keep the entrypoint lean and move depth into references or scripts."
|
|
},
|
|
{
|
|
"rank": 5,
|
|
"point": "Close the lifecycle loop",
|
|
"why": "Team-reused skills need visible ownership, review cadence, and regression evidence.",
|
|
"move": "Keep manifest, review viewer, and iteration directions aligned after each material change."
|
|
}
|
|
],
|
|
"reviewer_rule": "Reviewer should ask whether the skill's structure will keep producing the desired behavior after repeated real use."
|
|
} |