The claude-api skill, reviewed

What it tells the agent to do

1. Grep the project for other-provider markers; if any are found, stop and ask before writing Claude code
2. Detect the project language from its files, and ask the user if it is ambiguous
3. Pick the surface: a single call, a workflow, Tool Runner, or Managed Agents
4. Write code with the official SDK from documented bindings only, defaulting to claude-opus-5-5, adaptive thinking and streaming
5. If ANTHROPIC_API_KEY is unset, run `ant auth status` before asking the user for a key
6. For a bare subcommand, read the linked guide (e.g. model-migration.md) and execute it step by step

Lines worth reading

Straight from the file, with why each one matters.

1. line 5don't skip because it "looks like a one-liner

   The trigger argues with the agent in advance about skipping it. That is pushy, and it tells you how broadly this skill wants to load.

2. line 227that just means they were released after your training data cutoff - they are real models

   A neat fix for a real problem: it stops the model from 'correcting' new model IDs it doesn't recognize.

3. line 25don't reach for `requests`/`fetch` in a Python or TypeScript project just because it feels lighter

   A sharp, specific rule that blocks a common shortcut models take.

4. line 200Never downgrade for cost - that's the user's decision, not yours.

   A firm stance on model choice. It respects the user, but it also locks every default to the Opus tier.

5. line 62every run that exercises the model spends real money, so get the user's approval first

   A good guard: workflows that run the model ask before spending.

Watch out for

• Medium

  Tells the agent to look up local credentials and print a live OAuth access token instead of asking the user, which puts the token in the agent's context.

  line 237When you need to call the API and `ANTHROPIC_API_KEY` is unset, don't ask the user for a key.

• Medium

  The trigger is so broad that about 28k tokens load for almost any LLM-flavored task, even tiny ones, crowding out other context.

  line 5OR the task is LLM-shaped with provider unstated

• Low

  Hard-codes the Opus default and turns on server-side fallbacks by default, so generated code may cost more than the user expects.

  line 200This is non-negotiable.

• Low

  Fetches SDK repos and docs from the web at write time, so output depends on outside content and network access.

  line 27WebFetch the relevant SDK repo from `shared/live-sources.md` before writing code

The scorecard

• TriggerSays when to use it
  5/10

  Rich keywords, but over the 1,024-character limit and far too broad.

• FocusOne clear job
  5/10

  One domain, but a reference plus six workflows bundled together.

• LeanLight to load
  2/10

  About 105k characters and 28k tokens; huge tables that belong in reference files.

• ConcreteSteps you can follow
  8/10

  Exact model IDs, parameters, per-language snippets and commands throughout.

• SafeSafe to run
  6/10

  Prints auth tokens without asking; otherwise asks before spending money.

How it could be better

1. Cut the description under 1,024 characters

   It breaks the standard's cap, and the long trigger lists bury the key use.

   Add

   description: Write and debug code that calls the Claude API or Anthropic SDKs (model IDs, pricing, thinking, streaming, tool use, caching, Managed Agents, migration). Use when the request names Claude, Anthropic, Opus, Sonnet, Haiku or `anthropic` imports, or asks which LLM to use. Skip when the project uses OpenAI, Gemini or another provider.

   Copy

2. Move per-model notes into reference files

   Model sections, thinking tables and provider clients take most of the 28k tokens and are only needed now and then.

   Add

   ## Models
   Default: `<default model id>`. For per-model behavior, thinking rules and pricing, read `shared/models.md`. For cloud clients, read `shared/provider-clients.md`.

   Copy

3. Ask before reading or printing credentials

   Printing an OAuth token into the agent's context should be the user's call.

   Add

   If `ANTHROPIC_API_KEY` is unset, tell the user and offer to run `ant auth status`. Only run `ant auth print-credentials --access-token` after the user agrees.

   Copy

4. Narrow the provider-unstated trigger

   Loading the full reference for every 'summarize this text' task wastes context.

   Add

   Use this skill for LLM tasks with no provider named only when the user is writing code that will call a model API, not for one-off text work.

   Copy