topskills

Claude API guide

Best agent Skills for the Claude API

The Skills worth wiring into a Claude API integration are the document set (pdf, docx, xlsx, pptx) when your product generates real files for users, firecrawl or Valyu when it needs retrieval with citations, and Composio when it needs to reach many third-party services. 13 of 191 catalogued Skills are verified on the API surface — the narrowest list on this site, and the one where the choice is an architecture decision rather than a preference.

Updated August 26, 2026 · first published August 19, 2026

Install first

pdf

The clearest production case for a Skill on the API: your product needs to emit a real PDF, and a tested script beats generated code that must be correct every time for every customer.

Best when: A user-facing feature that produces or parses PDFs at volume.

Install model

How Skills install on Claude API

POST /v1/skills
  1. 1Choose between a pre-built skill_id and a custom upload — pre-built Skills are referenced by id with no hosting on your side.
  2. 2For a custom Skill, zip the folder and POST it to /v1/skills to get a skill_id back.
  3. 3Reference the skill_id in your Messages API request so the model loads it for that call.
  4. 4Version deliberately: pin the skill_id your integration was tested against rather than tracking a moving target in production.

On the API the Skill is part of your request configuration, not a file on a developer machine — which means it is also part of your deploy surface. Token cost is explicit and per-call: Skill metadata and any triggered instructions count against the request, so a Skill that fires on every call is a line item, not a free convenience.

Ranked

The shortlist

Ordered by what to install first for the common case, not by raw score. Every entry here carries verified Claude API compatibility — open the skill page for the composite score, licence, and full install commands.

SkillSourceScoreBest when
pdfOfficial52A user-facing feature that produces or parses PDFs at volume.
xlsxOfficial48Reporting and export features where the recipient will keep working in the file.
firecrawlVerified53Retrieval over many pages where you need structure out, not raw HTML.
composioVerified55A product feature that has to reach a long tail of SaaS tools.
valyuVerified48Answers that must be attributable — research, finance, anything a user will act on.
notion-api-skillVerified51Product features that read from or write into a customer Notion workspace.

See every skill verified on this agent, ranked by composite score, on the Claude API page.

Why these

Pick by pick

  1. 1
    pdf

    The clearest production case for a Skill on the API: your product needs to emit a real PDF, and a tested script beats generated code that must be correct every time for every customer.

  2. 2
    xlsx

    Real spreadsheets with working formulas, which is the difference between an export users trust and one they re-key by hand. Deterministic where improvised generation is not.

  3. 3
    firecrawl

    Hosted, JavaScript-rendering extraction that writes structured results rather than flooding the context window — the property that matters most when you are paying per token in production. Note it is action-taking and API-keyed, so it carries its own cost and failure modes.

  4. 4
    composio

    Apache-2.0, verified on all seven agent targets including the API. One integration surface instead of many hand-built connectors, which is usually the difference between shipping the feature and scheduling it.

  5. 5
    valyu

    Primary-source data with citations attached, read-only. On the API surface, where you cannot manually sanity-check every response, provenance you can show the user is worth more than fluency.

  6. 6
    notion-api-skill

    Notion's own Skill for reading and writing workspace content, which is the common integration when the customer's knowledge lives in Notion rather than your database.

Limits

What this agent cannot do

This is the narrowest list on the site: 13 of 191 Skills verified. That reflects what the surface is for — the API has no editor, no local repo, and no interactive session, so editor conventions and terminal workflows have nothing to attach to. Two API-specific constraints matter more than the count. First, cost is per call and explicit: a Skill that triggers on every request is a permanent token line item, so measure before you ship. Second, Skills become part of your deploy surface — pin the skill_id you tested, and treat a Skill upgrade as a release, not a settings change.

FAQ

Common questions

How do Skills work on the Claude API?

You either reference a pre-built skill_id or upload your own folder to /v1/skills and reference the id it returns. The Skill then loads for requests that include it, following the same three-tier model as the CLI: metadata always, instructions on trigger, bundled files on demand.

Do Skills increase my token usage?

Yes, and visibly. Metadata costs tokens on every call that includes the Skill, and a triggered Skill adds its instructions on top. That is usually still cheaper than the alternative — re-sending the same long prompt yourself — but it should be measured rather than assumed.

Can I use a community Skill in production?

You can, and you should read it first: a Skill is instructions your model will follow, so it is a supply-chain dependency. Check the license, pin a version, and re-read the diff when you upgrade. Prefer Anthropic and vendor-published Skills where a production guarantee matters.

Are Skills a substitute for tools or MCP on the API?

No, they are complementary. Tools and MCP servers let the model do things; a Skill tells it how and when. The document Skills are the clearest case — the Skill carries the method and the scripts, while your integration still owns file storage and delivery.

Other agents

Same question, different agent

New to Skills? What is a Claude skill · How to install · How to audit a skill first