Skip to main content
A skill tells an agent how to do one job repeatedly. SEC API does not prescribe a universal skill format, installer, scheduler, or secret store. Keep the SEC API portion portable: name the job, the retrieval path, the evidence to keep, and the point where the agent must stop.

​
Start with one filing question

This plain Markdown example can be adapted to the frontmatter and file location required by the agent product that will run it.
The intended output is an evidence-linked brief, not a general company assessment. A latest-filing rule can yield a different accession after a new filing arrives, so retain the selected filing identity in every run.

​
Choose REST or MCP at execution time

For a REST skill, verify paths and arguments in the API reference before publishing it. For an MCP skill, direct the client to use tools.search and tools.describe rather than embedding an assumed tool catalog. GET /mcp only reads public connection metadata; the skill still needs an authenticated MCP client to make a tool call. Do not write a ChatGPT skill that assumes an SEC API MCP OAuth connection: ChatGPT compatibility is not currently published or verified. API-key clients must keep the key in the process or desktop secret store that runs the client. Name the source fields the result needs, especially accessionNumber, filingDate, filingUrl, and requestId for filing work. Use one issuer, a defined filing family, a date rule, and a result limit where the route supports it.

​
Tell the agent when to stop

The skill should stop on authentication, authorization, validation, billing, quota, or unsupported-capability responses. It can use a bounded retry policy for transient failures, but it should preserve the request identifier and never retry by multiplying parallel calls. Before repeated work, inspect billing and limits and quote the representative operation when the workflow needs a current cost decision. An agent environment owns scheduling, state, and secret injection. Do not claim a SEC API skill creates those facilities or grants a key access it does not already have.

​
Test the skill with one real issuer

Run the skill once against a public issuer, inspect the raw response and final brief together, then add one test that fails when the expected filing identity or citation fields are absent.