I run a few side projects on LLM APIs, and the most annoying part of my week isn't debugging — it's looking things up. Pricing tiers split across input/output/thinking tokens. Rate limits buried three levels deep. Sample code that never matches the exact model I'm using. Four different portals (docs, console, model catalog, API reference) with four different layouts.
And none of that navigation muscle memory transfers between providers. Every platform organizes things differently, and they all ship new models and price changes every few weeks.
Here's the setup that made me stop browsing entirely.
The pattern: local docs, agent reads them
If you've used Context7, you know the idea — inject current docs into your AI assistant's context instead of letting it hallucinate from stale training data. That solves the open-source library side. But pricing, rate limits, and parameter schemas for a model platform don't live in any repo an MCP can pull.
Alibaba Cloud's Model Studio team shipped an Agent Skill applying the same idea to their platform: bailian-docs-llm-wiki. One install, and the platform's full documentation plus a structured snapshot of the model catalog become local files your agent reads directly.
Three layers:
| Layer | What it is | Used for |
|---|---|---|
models/ |
Structured catalog data, pulled from the console gateway API | Pricing, rate limits, context windows, sample code |
wiki/ |
Synthesized concept/comparison pages | Concepts, architecture choices |
raw/ |
Original official docs | API details, error codes |
The models/ layer is the differentiator: it's not scraped HTML — it's the same data endpoint that powers the catalog UI. QPM limits, tiered prices, parameter definitions, and official sample code are all structured fields.
Setup
npx skills add modelstudioai/skills
Node.js ≥ 18. Pick bailian-docs-llm-wiki at the prompt — zero config after that; questions touching the platform's docs, models, or pricing activate it automatically. Works in Claude Code, Qwen Code, Cursor — anything that supports Agent Skills. Queries read local files, so they cost nothing and need no API key. You'll want a key once you actually run models (free tier for new accounts): get an API key.
What it replaces, chore by chore
Pricing and rate limits. Ask "What does qwen3-max cost for input and output? What's the rate limit?" — tiered prices, QPM, and context window come back together, with the source file path cited. I checked the file the first time. Numbers matched.
Cross-family model filtering. "Which models support function calling with 128K+ context?" — the catalog UI can't filter like that; the skill greps models.jsonl, one model per line, exact field matching. The inverse question is the one I actually ask most — most of us route easy jobs to cheap models and save expensive ones for hard problems: "which budget model is good enough for this batch classification job?" Price is just another field to filter on. Terminal users can skip the agent entirely (examples from the skill's own docs):
jq -c 'select(.contextWindow>=1000000)' models.jsonl
grep '"function-calling"' models.jsonl | jq -c '{model,family,contextWindow}'
Sample code and parameters, in-editor. The samples field carries official examples in curl/Python/Node.js/Java; predictConfig carries parameter names, defaults, and ranges aligned with the platform playground. Ask in your coding agent, paste, adjust — and have the agent run the sample right there to verify before wiring it into your project. The whole loop stays inside the editor.
Error codes and concepts. Error codes get grepped from the raw official docs locally — no more fighting a docs site's search box. Concept questions ("RAG options on this platform?") hit pre-synthesized comparison pages.
Catching up on updates. When I hear a new model dropped, I re-run the install command to refresh the snapshot, then ask "what's new in the model list?" Note the boundary: it's a snapshot you refresh, not a push subscription. For me that's the right trade.
Why it doesn't make things up
Three constraints baked into the skill's rules:
- Answers must come from actual local files — no filling gaps from model memory for APIs, params, error codes, or prices.
- Every answer cites a file path you can open and verify.
- Conflict precedence:
models/(gateway data) >raw/(official docs) >wiki/(synthesized). Low-quality synthesized pages get bypassed automatically.
Not zero hallucination risk — but "verifiable by opening a file" beats "I think I found the right page."
Honest scope
Built for people working with Alibaba Cloud Model Studio or evaluating Qwen models. If your stack lives entirely on other providers, the skill deliberately refuses out-of-scope questions rather than improvising — correct behavior, but it won't help you there. Free tier if you want to poke at it: sign up.
What's your approach to keeping provider docs in your agent's context — Context7, llms.txt, something homegrown? Would love to compare notes.
Top comments (0)