meta-skill-creator

Meta-skill creator

Use this skill to turn repeatable workflows into skill packages that can run, be accepted, and be migrated. The product is not a "longer prompt," but a capability package with a very light default entry point, details loaded on demand, and reproducible validation methods.

Core contract

Start from the user's real workflow, not from a personally preferred template.
Before writing files, first prove that this is worth turning into a skill; do not force one-off tasks into a skill package.
SKILL.md Keep only the routing surface: triggers, first action, progressive loading, hard stops, and validation.
Put detailed research, product design, package plans, trigger evaluation, acceptance methods, and templates into references/、assets/、evals/、scripts/。
Skills that need to fill in ambiguous intent, have the user make multiple choices, or lock the MVP before batch generation must design Codex request_user_input / Claude Code AskUserQuestion and other host-native decision surfaces; Markdown choice cards can only be a degraded waiting interface and cannot impersonate real confirmation.
For images, videos, presentations, documents, reports, dashboards, or other rendered outputs, prioritize checking Image2 / host-native capabilities; MCP, scripts, SVG, and static previews are degraded or auxiliary routes and must have degradation evidence.
External materials can only be abstracted into quality principles; do not copy others' naming, page structure, visual systems, prompts, examples, or business phrasing.
When a candidate Skill will influence a public audience's understanding, preferences, or actions, or will generate multi-page visual content, the Skill must first research audience needs, public discussion, and comparable offerings on its own, then form an "opportunity research -> audience decision -> creative system." Do not turn traffic research into required user questions, and do not abstract a single platform's title/cover formulas into the general layer.
Every delivery must close the loop: record runtime evidence, review findings,writeback / proposal / none-with-reason / blocked decisions, and write reusable learnings back to the corresponding reference, template, validator, eval, or example.

Capability layering

A complete package is a capability library, not a fixed template for all candidate Skills. Choose intensity according to task risk and delivery stage:

core: All candidates must clarify goals, triggers, inputs, outputs, execution step contracts, boundaries, and minimum validation. The goal must state who gets what final product under what pressure and how it will be accepted; each step must have inputs, actions, observable outputs, conditions for proceeding to the next step, and failure routes; boundaries must cover non-goals, permissions, sources of truth, side effects, stopping, and degradation.
conditional: Only load deep research, content opportunities, creative decision boards, multimodal, or interactive MVP modules when the domain is unfamiliar, facts may change, outputs are visual, the audience is public, multiple rounds of decision-making are needed, or host tools are required.
release: Only when claiming distributable, sellable, migratable, or clearly superior to the baseline, enforce clean-session baselines, host proof, human acceptance, and a complete closed loop.

Do not require every candidate Skill to fill in a template just because it exists in the complete package; nor should you delete capabilities that will be needed when the hit conditions are met later just because the current task is simple.

Progressive loading

Read only the files needed for the current stage:

Current stage	Read	When to use
Intent and domain research	`references/intent-domain-research.md`、`assets/domain-research-brief-template.md`	All candidates first write a minimal domain research brief; upgrade to deep research only when the domain is unfamiliar, current facts will change, the user requests deep research, or public compatibility/security/commercial-readiness claims are involved. If evidence is still insufficient after retrieval, mark `research-needed`. Content marketing or public communication Skills then load content opportunity research.
Surface and artifact chain	`references/experience-surface-model.md`	The Skill must produce files, media, screenshots, presentations, reports, dashboards, or other visible artifacts.
Productization design	`references/product-design.md`	When it is necessary to decide on multiple artifacts, the user journey, a 3-minute visible result, the first public surface, or commercial delivery. A simple single workflow is not mandatory.
Runtime contract	`references/skill-contract.md`	Use before writing or modifying a candidate skill package.
Source abstraction boundary	`references/source-abstraction-boundary.md`	When using external cases, internal samples, or competitor materials, abstract first, then rewrite.
Package plan	`assets/package-plan-template.md`	Map before editing `SKILL.md`, reference files, assets, scripts, evaluations, and examples.
Multimodal/tool route	`references/multimodal-tooling.md`、`assets/multimodal-prompt-brief-template.md`	The Skill depends on Image2, host media capabilities, MCP, a local renderer, or scripts; first record the local capability inventory and multimodal brief. For multi-page visual artifacts, also write the role, visuals, text area, forbidden occlusion area, and page continuity for each page.
Interactive decision-making and MVP gate	`references/interactive-mvp-product-skill.md`	The Skill must pop up multiple decisions under ambiguous intent, lock style/cover/first screen/MVP before batch production, or reuse the confirmation process across Codex and Claude Code hosts.
Evaluation design	`references/evaluation-method.md`、`evals/trigger-eval.json`	`core` Design at least positive and negative trigger examples, two real tasks, and one boundary task.
Release-level runtime proof	`assets/trigger-run-record-template.json`、`assets/host-event-anchor-template.json`、`assets/host-attestation-template.json`、`assets/reviewer-attestation-template.json`	Only at the `release` layer, design real with/without clean-session, out-of-directory run/manual anchor, and baseline comparison.
Release gate	`references/release-gate.md`、`assets/acceptance-run-template.md`	Before claiming it can be distributed, migrated, or delivered to others for use.
Closed-loop governance	`references/closed-loop-governance.md`、`assets/loop-run-record-template.md`	During acceptance, failure, retrospective, drift, release preparation, or when the user requests a "closed loop"; runtime records and write-back decisions must be stated.

assets/ in .md is a work template for people to fill in,.json is structured data for scripts or validation to read; read it only when creating the corresponding artifact.scripts/ runs only during validation, conversion, or packaging; do not paste script contents into the conversation context.

Workflow

Classification: determine whether it is new-skill、refactor-skill、evaluate-skill、package-plan, or not-a-skill。
Fetch and research: first do the minimal evidence scan sufficient to verify the goal, native artifacts, and failure modes; only when the conditional conditions are hit, upgrade and execute the references/intent-domain-research.md Abstract Deep Research Protocol. Collect only evidence that will change package decisions, artifact surfaces, tool routes, or acceptance criteria. For public communication/marketing Skills, the Skill itself researches audience needs, scenario terms, similar content, public discussions, and content gaps, and does not ask the user "where does the traffic come from".
Design: determine outcomes, trigger boundaries, inputs and outputs, local capability route, interactive decision surface, MVP lock gate, package structure, and public failure modes. Multi-page visual artifacts must have a Creative Decision Board that can explain what each page draws, where text goes, and why it is arranged that way.
Build: create or modify only the minimal set of files that the current goal must touch.
Verify:core layer runs structural checks, positive and negative trigger examples, two real tasks, and one boundary task;release layer then uses the cross-platform Python entry point to create the pending acceptance directory, where an external host separately runs clean sessions without/with the Skill loaded, and leaves host event/attestation and human reviewer attestation outside the run folder. Report what structure_status=artifact_consistent、runtime_status=verified、human_status=verified each proves respectively; neither layer can substitute for the other.
Close the loop: fill in the run record; decide writeback、proposal、none-with-reason or blocked; only reusable learnings are written back to the rules; a chat summary must not be passed off as a closed loop.

Package structure

Available modules (not the default file list for every candidate Skill):

SKILL.md: lightweight entry point, responsible for triggering and the execution contract.
references/: domain rules, product model, capability route, source abstraction, evaluation methods, release gates.
assets/: fillable templates, worksheets, structured data, or prompt briefs.
scripts/: deterministic validation, transformation, rendering, or package checks.
evals/: trigger evaluation and regression cases.
examples/: small but real input/output examples; do not write private history, internal scores, or fake run evidence.

Add new files only when they prevent a concrete failure or make the Skill easier to run through. Do not add reference files just to store research notes.

Hard stop

Stop first and explain the blocker when encountering the following:

The task is just a one-off prompt, not a repeatable workflow.
The goal only states a feature or file name, without users, pressure scenarios, final artifacts, and observable success criteria.
The execution steps contain only empty actions such as "analyze, optimize, process", without inputs, actions, outputs, transition conditions, or failure routes.
The boundaries do not state non-goals, sources of truth, permissions/side effects, stop conditions, and degradation behavior.
The domain surface is guessed, without support from user materials, local evidence, or current source evidence.
A multimodal Skill has no local capability inventory and output evidence route.
It requires multiple rounds of user decisions or MVP confirmation, but has no native selection tool, degraded waiting interface, confirmation status field, and batch generation stop rule.
The package depends on generated files, but there is no way to prove that files, screenshots, media, or rendering results exist.
The candidate package wants to claim public-ready, sellable, transferable, or better than baseline, but acceptance / baseline only has the pending template generated by the preparation script, or only has in-run logs/local hashes, without two real sessions that are different after normalization, out-of-directory host event/attestation, per-dimension human scoring, and out-of-directory reviewer attestation.
The trigger scope overlaps with another active skill package, but there is no routing boundary.
The design would expose private evaluation history, internal retrospectives, implementation scars, or make it obvious that it imitates some external project.

Verification

The following commands accept meta-skill-creator 's own complete contract, not fixed commands for all candidate Skills:

python scripts/check_meta_skill_package.py .
python scripts/check_closed_loop.py .
python scripts/test_acceptance_runs.py
python scripts/prepare_acceptance_run.py --mode acceptance --input-file <raw-input-file>
python scripts/check_acceptance_runs.py <completed-run-dir>

prepare_acceptance_run.py Only prepare templates; do not call models, pre-fill scores, or produce PASS; after preparation is complete, check_acceptance_runs.py failure is expected behavior. The local hash only proves artifact consistency; only after real clean-session output, out-of-directory host event/attestation, per-dimension human scoring against a frozen rubric, and out-of-directory reviewer attestation are all bound can runtime/human become verified, and only then is Ready allowed to pass. Per-dimension evidence_refs must point to lines or line ranges that exist and are non-empty in both real outputs; wrapping placeholders, AI/automated reviewers, and forged or out-of-bounds line numbers must all fail. If the current system does not have the python command, you can use py -3; on Linux/macOS prefer python3. The shell wrapper will reject interpreters below 3.10.

The candidate Skill should run its own scripts and evaluations. Only when entering the release layer are the full clean-session, host attestation, and human acceptance gates applied.

Final report

Report in Chinese: which files were changed, package decisions, trigger boundaries, progressive loading diagram, validation commands, proof tiers, closed-loop decisions, and remaining risks. Provide a small before-and-after comparison or output snippet so the user can judge whether this skill package is truly easier to use.