Aliyun Skill Creator

Category: tool

Alibaba Cloud Skill Creator

Repository-specific skill engineering workflow for alicloud-skills.

Use this skill when

• Creating a new skill under skills/**.
• Importing an external skill and adapting it to this repository.
• Updating skill trigger quality (name and description in frontmatter).
• Adding or fixing smoke tests under tests/**.
• Running structured benchmark loops before merge.

Do not use this skill when

• The user only needs to execute an existing product skill.
• The task is purely application code under apps/ with no skill changes.

Repository constraints (must enforce)

• Skills live under skills/<domain>/<subdomain>/<skill-name>/.
• Skill folder names use kebab-case and should start with alicloud-.
• Every skill must include SKILL.md frontmatter with name and description.
• skills/**/SKILL.md content must stay English-only.
• Smoke tests must be in tests/<domain>/<subdomain>/<skill-name>-test/SKILL.md.
• Generated evidence goes to output/<skill-or-test-skill>/ only.
• If skill inventory changes, refresh README index with scripts/update_skill_index.sh.
• Every create/refresh pass MUST start by re-fetching the upstream official docs — do not edit a skill from memory or from the previous draft alone. See "Fetching latest Alibaba Cloud documentation".

Standard deliverable layout

skills/<domain>/<subdomain>/<skill-name>/
├── SKILL.md
├── agents/openai.yaml
├── references/
│   └── sources.md
└── scripts/ (optional)

tests/<domain>/<subdomain>/<skill-name>-test/
└── SKILL.md

Workflow

1. Capture intent

• Confirm domain/subdomain and target skill name.
• Confirm whether this is new creation, migration, or refactor.
• Confirm expected outputs and success criteria.
• Re-fetch the upstream documentation now, even for a refresh — see "Fetching latest Alibaba Cloud documentation". Diff what you find against the existing skill before changing anything.

2. Implement skill changes

• For new skills: scaffold structure and draft SKILL.md + agents/openai.yaml.
• For migration from external repo: copy full source tree first, then adapt.
• Keep adaptation minimal but explicit:
  • Replace environment-specific instructions that do not match this repo.
  • Add repository validation and output discipline sections.
  • Keep reusable bundled resources (scripts/, references/, assets/).
• Always ground the skill in the latest official Alibaba Cloud documentation (see "Fetching latest Alibaba Cloud documentation" below) and record every authoritative URL in references/sources.md.

3. Add smoke test

• Create or update tests/**/<skill-name>-test/SKILL.md.
• Keep it minimal, reproducible, and low-risk.
• Include exact pass criteria and evidence location.

4. Validate locally

Run script compile validation for the skill:

python3 tests/common/compile_skill_scripts.py \
  --skill-path skills/<domain>/<subdomain>/<skill-name> \
  --output output/<skill-name>-test/compile-check.json

Refresh skill index when inventory changed:

scripts/update_skill_index.sh

Confirm index presence:

rg -n "<skill-name>" README.md README.zh-CN.md README.zh-TW.md

Optional broader checks:

make test
make build-cli

5. Benchmark loop (optional, for major skills)

If the user asks for quantitative skill evaluation, reuse bundled tooling:

• scripts/run_eval.py
• scripts/aggregate_benchmark.py
• eval-viewer/generate_review.py

Prefer placing benchmark artifacts in a sibling workspace directory and keep per-iteration outputs.

Definition of done

• Skill path and naming follow repository conventions.
• Frontmatter is complete and trigger description is explicit.
• Test skill exists and has objective pass criteria.
• Validation artifacts are saved under output/.
• README skill index is refreshed if inventory changed.
• references/sources.md lists every upstream URL consulted in this pass, with last-checked date in the note.
• Model ids, parameter names, quotas, and pricing in SKILL.md cite a URL fetched in this pass — none copied forward unverified.

Fetching latest Alibaba Cloud documentation

Hard rule: every create or refresh pass starts here. Re-fetch the upstream doc tree first, diff it against the existing skill, then write. Do not trust model memory and do not trust the prior draft — Alibaba Cloud rotates model ids, parameter names, and pricing on a weekly cadence.

Order of authority (stop at the first source that fully answers the question):

1. Help Center (help.aliyun.com/zh/...) — primary source

The Chinese path (/zh/) ships first and is the most accurate. The English path (/en/) often lags by weeks; only link to it after confirming parity.

Fetch with WebFetch:

WebFetch(url="https://help.aliyun.com/zh/<product-slug>/<api-reference-slug>",
         prompt="Extract endpoint, parameters, response schema, model ids, and limits.")

If the response is a redirect notice, re-issue WebFetch against the redirect URL.

2. Bailian / Model Studio (/zh/model-studio/...) — entry portal

For any DashScope-backed skill (LLM, image, video, audio, embedding, rerank, agent, application), the canonical entry is:

https://help.aliyun.com/zh/model-studio/models

From this page, walk into the topic sub-trees below. Always check newly-released-models first — that page is the source of truth for which model ids are GA right now and which were silently retired.

If a slug is unknown, GET the entry page first and let the sidebar tell you the current path — slugs occasionally change.

DashScope runtime base URLs:

Mainland:   https://dashscope.aliyuncs.com/api/v1/...
Singapore:  https://dashscope-intl.aliyuncs.com/api/v1/...

3. OpenAPI Portal metadata (api.aliyun.com) — machine-readable spec

Use this for non-DashScope products (ECS, OSS, RDS, SLS, ESA, …) when you need exact request/response field types or a complete API list.

# All products (one-time index, EN labels)
https://api.aliyun.com/meta/v1/products.json?language=EN_US

# All APIs for a product/version
https://api.aliyun.com/meta/v1/products/<Product>/versions/<YYYY-MM-DD>/api-docs.json

# Single API definition (full schema)
https://api.aliyun.com/meta/v1/products/<Product>/versions/<YYYY-MM-DD>/apis/<ApiName>/api.json

Repository helpers already wrap these endpoints — reuse them rather than writing one-off scrapers:

python3 scripts/products_from_openapi_meta.py        # refresh product index
python3 scripts/apis_from_openapi_meta.py            # fetch APIs per product (env-driven)

Browsable portal: https://api.aliyun.com/product/<Product> — for human inspection and to confirm a <Product> code before scripting.

4. Official SDKs and CLIs

5. Recording sources

Every skill MUST list the URLs it actually consulted in this pass, in references/sources.md:

# Sources

- [<Human title>](<URL>) — what this covers (last checked YYYY-MM-DD)

Include at minimum: API reference, model/version notice (newly-released-models for Model Studio), pricing (if relevant), and SDK home page. Date-stamp each entry — that is how the next maintainer knows what is stale.

Red flags — STOP and re-fetch

• Editing the skill without opening any upstream URL in this session.
• Copying model ids, parameter names, or quotas forward from the existing draft.
• Linking only to /en/ pages without checking /zh/.
• Saving raw HTML/JSON dumps under references/ instead of summarizing into SKILL.md.
• Writing curl loops when scripts/products_from_openapi_meta.py / scripts/apis_from_openapi_meta.py already cover the call.

References

• references/schemas.md
• references/sources.md