claude-api skill review: is Anthropic's Claude API skill worth using?
The claude-api skill is Anthropic's reference skill for Claude API and SDK code in eight languages. It comes bundled with Claude Code. We rate it Recommended with caveats, 72/100, at Moderate confidence. Its current model IDs and prices check out. Retired models, loader bugs and a heavy entry file hold it back.
- Skill
- Claude API (anthropics/skills, skills/claude-api)
- Verdict
- Recommended with caveats
- Score
- 72/100
- Confidence
- Moderate
- Safety grade
- B
- Best for
- Developers in Claude Code who write or migrate Anthropic SDK code and want current model IDs, prices and API shapes
- Poor fit
- Strict loaders such as Pi or Copilot CLI; context-tight sessions where a ~25K-token entry file is a real cost
- Reviewed version
- no version · commit 8a1541c
- Review date
- 2026-09-30
- Works in (verified)
- Claude Code
- Works in (claimed)
- Claude.ai, Claude Agent SDK
- Status
- Updated since review: 3 commits from 2026-10-05 to 2026-10-09 (Haiku 5.5, Managed Agents)
Get the skill: claude-api folder on GitHub · SKILL.md at the reviewed commit · skills.sh listing
What is the claude-api skill?
The claude-api skill is Anthropic's reference skill for writing Claude API and SDK code. It covers the Messages API and Managed Agents in eight languages. It ships bundled with Claude Code and in anthropics/skills.
An Agent Skill is a folder of instructions, led by a SKILL.md file. An agent loads it when a task matches. Our explainer on what Agent Skills are covers how that loading works. The claude-api skill is one of these Claude skills, published by Anthropic in its anthropics/skills repository. Its job is narrow. It gives a coding agent current model IDs, parameters, prices and migration rules. The aim is generated code that runs on the first try.
The skill exists because models drift. A model's training memory holds old model names and old API shapes. SKILL.md tells the agent to trust the skill's tables over that memory. That rule is the skill's main strength and the source of its worst failure.
The Messages API is the main Claude API endpoint for sending prompts and receiving replies. Claude Managed Agents is a beta service for hosted agent sessions. Anthropic's doc page says Managed Agents runs only on the Claude API and Claude Platform on AWS.
| Fact | Value |
|---|---|
| Publisher | Anthropic |
| Path | anthropics/skills, skills/claude-api |
| Files at review | 81: SKILL.md (about 102 KB), 63 reference docs, 4 scripts |
| Surfaces covered | Messages API; Claude Managed Agents (beta) |
| Licence | Apache 2.0 (LICENSE.txt, fetched 2026-10-21) |
| Distribution | Bundled with Claude Code; installable from anthropics/skills |
| Version | None; reviewed at commit 8a1541c |
The skill covers eight languages, as listed on Anthropic's doc page:
| Language | What the skill holds |
|---|---|
| Python | SDK reference docs |
| TypeScript | SDK reference docs |
| C# | SDK reference docs |
| Go | SDK reference docs |
| Java | SDK reference docs |
| PHP | SDK reference docs |
| Ruby | SDK reference docs |
| cURL | Raw HTTP examples, with no tool runner |
A shared folder holds the cross-language material. That material includes model tables, pricing, prompt caching, batches and the model migration guide.
Is the claude-api skill the same as the Skills API?
No, the claude-api skill is not the Skills API. The skill teaches Claude to write API code. The Skills API is a set of endpoints for uploading and running any Agent Skill through the API.
| Aspect | claude-api skill | Skills API |
|---|---|---|
| What it is | Reference material an agent reads | The /v1/skills endpoints on the Claude API |
| Purpose | Write correct Claude API and SDK code | Upload, version and run any skill in API requests |
| Where it runs | Bundled in Claude Code; installable elsewhere | Messages API requests with the code execution tool |
| Limits | Description cap of 1,024 characters applies | Upload under 30 MB; up to 20 skills per request |
Anthropic's overview notes that skills do not sync across surfaces. A skill you upload through the Skills API is separate from the copy in Claude Code. Both read the same file format, explained in our guide on how SKILL.md frontmatter works.
Is the claude-api skill worth using?
Yes, the claude-api skill is worth using for current-model work in Claude Code. It scores 72/100, Recommended with caveats, at Moderate confidence. Its model IDs and prices were correct at review, but it loads heavily and lists retired models.
The strength is first-party upkeep. All 13 current model IDs match Anthropic's models overview. All 12 rows of the SKILL.md price table match the pricing page. The skill lists cache reads at $0.25 for Fable 5.1 and $0.20 for Opus 5.5. It lists Opus 5.5 fast mode at $8 input and $40 output. Fable 5.1 batch sits at $5 and $25. Each of those matches the pricing page. Opus 5.5 and Sonnet 5.5 arrived in one pull request within days of launch.
The weaknesses sit at the edges. Four retired models still appear in models.md as active or deprecated. They are Opus 4.1, Opus 4, Sonnet 4 and Haiku 3. The description runs past the 1,024-character cap, so strict loaders reject the skill. Pi, a coding agent with a strict skill loader, is one. Its provider grep treats the word "coherent" as a competitor's name. The entry file costs about 25K tokens on every load.
The score leans on the worst failure. An agent that trusts the model list can ship code that calls a retired ID. That risk falls on requests for older models. Our page on how we score and cap skills defines each measure. The list of skills ranked by verdict shows where this skill sits among the rest.
| Measure | Score | Reason |
|---|---|---|
| Effectiveness | 78 | High file quality; no with/without benchmark |
| Reliability | 62 | Retired-model rows, a misfiring grep and contested batch advice |
| Safety | 85 | No telemetry; the API probe is opt-in and never prints keys |
| Efficiency | 50 | About 25K tokens per load; past incidents reached 320K |
| Compatibility | 55 | Works in Claude Code; Pi and Copilot CLI fail; five file collisions |
| Maintainability | 72 | Frequent first-party updates; community fixes wait for weeks |
| Setup friction | 85 | Bundled in Claude Code; one command elsewhere |
| Model ID and lifecycle accuracy | 68 | 13 current IDs match; four retired models still listed |
| Pricing accuracy | 90 | 12 price rows, cache, batch and fast-mode rates match |
| Activation correctness | 55 | The grep matches "coherent"; prose model names trigger loads |
| Loader compatibility | 40 | Description over the 1,024-character cap; Tessl marks it invalid |
| Instruction weight vs value | 50 | Entry file still about 25K tokens after the harness fix |
| Guidance currency | 85 | New models land within days; one caching caveat missing |
No independent with-and-without benchmark exists. Tessl, a skill registry that validates skills, ran no evals for this skill. Stale lifecycle rows are drift, and the skill invents nothing.
The best fit is developers in Claude Code who write or migrate Anthropic SDK code. They get current model IDs, prices and API shapes. Poor fit covers three groups:
- Pi, Copilot CLI or other strict loaders that enforce the 1,024-character description cap.
- Repos whose text holds words like "coherent", which trip the unanchored SKIP grep.
- Context-tight sessions or subagents where a ~25K-token entry file is a real cost.
How many tokens does the claude-api skill use?
The claude-api skill loads 25,431 tokens of SKILL.md per use, our measured file size per load. Reference files add up to 274,115 more on demand. A typical coding task costs about 32,000 input tokens by our estimate.
| What | Tokens |
|---|---|
| SKILL.md, loaded each time the skill runs | 25,431 |
| Linked reference files, loaded on demand | 274,115 |
| Whole folder | 299,546 |
| Typical task, our estimate | 32,000 input, 3,000 output; 15 runs a week |
| Typical task, including the skill file | 58,431 input, 3,000 output |
The typical task is adding a streaming tool-use call to an existing Python service. The 58,431 figure includes the skill's own load, so the two input figures differ.
Measured incidents show the earlier cost. Issue #1487 reports about 156K tokens injected, which ended a default-context session. Issue #1486 reports 320K tokens, about $3.50 in API cost, just to load the skill. Issue #1439 measured 246K against 58K total tokens on one A/B agent task. The New Stack is a tech news site. It reports that Claude Code v2.1.234 cut the load from about 200K to about 25K tokens. Claude Code now reads the reference docs on demand.
Two hidden costs remain. Subagents pay the entry file again on each load. The trigger fires on prose mentions of model names, so the skill loads during plain discussion.
What changed since our review?
Anthropic changed the claude-api skill three times after our review, between 2026-10-05 and 2026-10-09. The updates add Claude Haiku 5.5 and Managed Agents material. The retired-model, description and grep defects remain.
The three commits:
- 2026-10-05, commit 683bc88 (#1962): managed-agents-onboard starts from a quickstart name or URL.
- 2026-10-09, commit 9d63080 (#1991): Claude Haiku 5.5 becomes the current Haiku, with Managed Agents and preserved-thinking fixes.
- 2026-10-09, commit dbd4588 (#1995): dynamic workflows and four workflow quickstarts.
Together the three commits changed 63 files in skills/claude-api and added 15 of them. They added 2,712 lines and removed 409. SKILL.md is now 105,146 bytes.
Three defects remain at dbd4588. The description is still 1,077 characters (#1622). The unanchored cohere grep is still there (#1624). Opus 4.1, Sonnet 4, Opus 4 and Haiku 3 still sit under "Deprecated" in models.md (#1603). PR #1607, the retired-model fix, was closed unmerged on 2026-10-05. Our verdict covers commit 8a1541c and stands until we review the new commits.
Is the claude-api skill safe to install?
Yes, the claude-api skill is safe to install: it carries safety grade B. It has no telemetry or hidden endpoints. One optional probe calls the Claude API with your own key after explicit confirmation.
The skill is mostly Markdown reference files plus four helper scripts. It has network access. Our guide on how we grade skill safety lists what we check. Our review of the files at commit 8a1541c found these eight facts:
- Files: 81 in total, with SKILL.md at about 102 KB. The 63 reference docs span eight language folders and shared/.
- Probe: drop_block_probe.py replays captured requests to api.anthropic.com, or to ANTHROPIC_BASE_URL.
- Credentials: the probe uses ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN, plans only without --yes, and never prints credentials.
- Local tool: prefix_diff.py diffs prompt prefixes locally, with no network.
- Reports: two Node eval-report scripts use no-follow file handling and formula-injection escaping, with no network calls found.
- Extras: no hooks, MCP config or telemetry exist.
- Live docs: the skill tells the agent to WebFetch about 70 platform.claude.com URLs, and all returned 200 when checked.
- Boundaries: the skill tells the agent to stop and ask before editing non-Anthropic provider files.
Gen Agent Trust Hub, Socket and Snyk audits on skills.sh, a public directory of Agent Skills, all report Pass.
What does the community report about the claude-api skill?
The community reports accurate data and heavy loading for the claude-api skill. Checked prices and current model IDs match Anthropic's docs. GitHub issues document retired models listed as usable, loader failures and sessions exhausted by context.
These are the 18 dated reports we weighed, 10 of them strong. Thirteen are negative, one is neutral and four are positive. Most come from GitHub issues and pull requests on anthropics/skills.
| Claim | Source | Month | Sentiment | Strength |
|---|---|---|---|---|
| Five model IDs in models.md return 404, including Opus 4.1 under "still active" | Issue #1603 | 2026-08 | Negative | Strong |
| Community PR moves four retired IDs to the retired table; reviewed, unmerged | PR #1607 | 2026-08 | Negative | Weak |
| Unanchored SKIP grep matches "coherent" and disables the skill in unrelated repos | Issue #1624 | 2026-08 | Negative | Strong |
| Description of 1,077 characters exceeds Pi's 1,024 cap; skill dropped at load | Issue #1622 | 2026-08 | Negative | Strong |
| Description truncated in listings; a later comment says Copilot CLI refuses it | Issue #881 | 2026-04 | Negative | Weak |
| Validation marks it invalid: 1,068-character description, 573-line SKILL.md | Tessl quality page | 2026-09 | Negative | Strong |
| Invoking the skill injected about 156K tokens and killed a default-context session | Issue #1487 | 2026-07 | Negative | Strong |
| Loading the skill consumed 320K tokens, about $3.50 in API cost | Issue #1486 | 2026-07 | Negative | Weak |
| A/B on one agent task: 246K against 58K total tokens with the skill present | Issue #1439 | 2026-07 | Negative | Strong |
| Claude Code v2.1.234 cut the load from about 200K to about 25K tokens | The New Stack | 2026-08 | Positive | Strong |
| Messages API cache priming gave a batch 0 cached tokens; skill advice conflicts | Issue #1475 | 2026-07 | Negative | Strong |
| Router file about 9,800 words; no eval suite for routing or code quality | Issue #1428 | 2026-07 | Negative | Weak |
| Linter found five same-name file collisions in the language folders | Issue #920 | 2026-04 | Negative | Weak |
| Four dead doc URLs reported in live-sources.md; all resolve at the pinned commit | Issue #1698 | 2026-08 | Neutral | Weak |
| Maintainer corrected two wrong refusal-billing claims and linked the docs | PR #1825 | 2026-09 | Positive | Weak |
| Opus 5.5 default and Sonnet 5.5 added with pricing within days of launch | PR #1930 | 2026-09 | Positive | Weak |
| Official prices for current models match the skill's tables, cache and fast-mode rates | Claude pricing docs | 2026-09 | Positive | Strong |
| Deprecations page confirms Opus 4.1, Opus 4, Sonnet 4 and Haiku 3 are retired | Model deprecations docs | 2026-09 | Negative | Strong |
The two description figures differ by source. Issue #1622 counts 1,077 characters, and Tessl counts 1,068. Both exceed the 1,024-character cap.
What are the claude-api skill's known failures?
The claude-api skill has four main known failures at commit 8a1541c. Retired models still read as usable, and the description breaks strict loaders. A loose grep disables it in unrelated repos, and the entry file is large.
- Retired models: models.md lists Opus 4.1, Opus 4, Sonnet 4 and Haiku 3 as active or deprecated. Issue #1603 reproduced five 404 responses.
- Long description: the frontmatter description exceeds 1,024 characters, so Pi drops the skill at load (#1622).
- Loose grep: the unanchored cohere pattern matches "coherent", so the skill skips itself in unrelated repos (#1624).
- Heavy entry file: SKILL.md costs about 25K tokens per load, and every subagent pays it again.
- Contested caching advice: a measured test gave a batch 0 cached tokens after Messages API priming (#1475).
- Name collisions: a linter found five same-name files across language folders, which break flat installs (#920).
The worst failure joins two of these. The agent trusts the skill's "only use IDs listed here" rule and ships code pointing at a retired model. Or the skill silently skips or overflows the context, and the agent falls back to stale memory. Our guide on why a skill does not trigger covers the skip case. Check legacy model IDs against Anthropic's deprecations page before you ship.
Which agents does the claude-api skill work with?
The claude-api skill is verified to work in Claude Code, where it ships bundled. Anthropic says it installs in any agent that supports Agent Skills. Pi rejects it, and GitHub Copilot CLI fails to load it.
| Agent | Status | Evidence |
|---|---|---|
| Claude Code | Verified; bundled | Anthropic docs; issues #1487 and #1439 |
| Claude.ai | Claimed by Anthropic | Publisher listing |
| Claude Agent SDK | Claimed by Anthropic | Publisher listing |
| Pi | Fails: description over the 1,024 cap | Issue #1622 |
| GitHub Copilot CLI | Fails to load | Issue #881 comment |
| Other Agent Skills clients | Unknown; no published run | None |
Compatibility scores 55. The skill needs only WebFetch, and only for live docs. It needs no MCP server. Browse Claude Code skills ranked by verdict for other options in the agent where this one works best.
How do you install the claude-api skill?
You install nothing in Claude Code, because the claude-api skill ships bundled there. Elsewhere, you install it with npx skills add, or add the anthropics/skills marketplace and install the claude-api plugin.
In Claude Code
The skill comes with Claude Code, so there is nothing to install. Call it directly by typing /claude-api, alone or with a subcommand. Anthropic's Claude Code docs explain how to turn bundled skills off in settings. Our Claude Code skills setup guide covers skill folders, settings and reloading.
Outside Claude Code
Two commands from Anthropic's doc page install the skill in other environments. Use the skills CLI:
npx skills add https://github.com/anthropics/skills --skill claude-apiOr add Anthropic's plugin marketplace and install the plugin:
/plugin marketplace add anthropics/skills
/plugin install claude-api@anthropic-agent-skillsThe marketplace name is anthropic-agent-skills. Strict loaders such as Pi reject the skill until the description fits the cap.
How do you use the claude-api skill?
You use the claude-api skill by asking Claude Code to build or fix Claude API code. It triggers when code imports an Anthropic SDK. Type /claude-api with a subcommand, such as migrate, to call it directly.
Anthropic's doc page sets the activation rules. The skill loads automatically when code imports anthropic or @anthropic-ai/sdk. It stays out when the code uses another AI provider's SDK. Its SKIP check is the grep that misfires on "coherent".
The evidence supports six jobs, ranked by how often they show up:
- Write new Messages API code with the right model ID: streaming, tool use, structured outputs (primary, daily).
- Migrate code between model generations, including thinking, tool_choice and sampling changes (primary, per model launch).
- Look up current prices, cache and batch rates to estimate cost (secondary, weekly).
- Set up prompt caching and batch workloads (secondary, occasional).
- Build Managed Agents sessions, environments, webhooks and memory stores (claimed-only).
- Build and hill-climb evals for Claude-backed features (claimed-only).
The two claimed-only jobs rest on Anthropic's own pull requests, and no independent results for them are published.
Test it yourself before you rely on it:
- Open a fresh Claude Code session.
- Ask for a Python streaming call with one tool on "opus".
- Run the code with your own key.
- Check that the model ID and the thinking parameters work.
What are the alternatives to the claude-api skill?
No skill we have reviewed competes with the claude-api skill. The closest alternatives are Anthropic's own SDK repositories and the Claude cookbooks. They give examples, but not current model and pricing rules for an agent.
| Alternative | What it is | Status |
|---|---|---|
| anthropics/anthropic-sdk-python | The Python SDK repository, with examples and no agent rules | Unevaluated |
| anthropics/claude-cookbooks | Recipe notebooks for common Claude API tasks | Unevaluated |
Both are plain repositories, so an agent reads them only when you point it there. MCP Builder, which we rate Conditional at 55, builds MCP servers, a different job. Anthropic's courses on Skilljar and Claude Academy teach the Claude API and Agent Skills to people. They are learning resources, and no agent loads them. Our full verdict ranking lists every skill we have reviewed.
Where is the claude-api skill on GitHub?
The claude-api skill lives in the anthropics/skills GitHub repository at skills/claude-api. Our reviewed commit holds 81 files across eight language folders and shared references. It is Apache 2.0 licensed, unlike the document skills.
- Repository: anthropics/skills, folder skills/claude-api, with 15 contributors on GitHub.
- Popularity: 179,227 stars on the repository at review; 68.3K installs of this skill on skills.sh.
- Copies: 1,182 exact copies of the SKILL.md exist in other repositories.
- Reviewed commit: 8a1541c, checked on 2026-09-30.
- Current folder head: dbd4588, dated 2026-10-09.
- Releases: none; the skill has no version number.
The LICENSE.txt file in the folder, fetched on 2026-10-21, says Apache 2.0.
Common questions
Which subcommands does the claude-api skill have?
Anthropic's Claude Code docs list nine: migrate, upgrade, managed-agents-onboard, managed-agents-secure, prompt-audit, cost-optimize, build-eval, hillclimb and preserved-thinking-migration. That list was checked on 2026-10-21.
Does the claude-api skill audit prompts?
Yes, Anthropic lists a /claude-api prompt-audit subcommand. We have not tested it, so its behaviour is claimed by Anthropic.
Where are the Agent Skills docs?
Anthropic's Agent Skills overview on the Claude Platform docs is the official reference. Our explainer on Agent Skills explains the format in plain terms.
Does the claude-api skill work outside Claude Code?
Yes, Anthropic says it installs in tools that accept its frontmatter, through npx skills add or the plugin marketplace. Pi and Copilot CLI fail to load it.