AI agents

Ziggeo publishes drop-in instructions for coding agents. Install one file and your agent knows how to record, upload, transcode and play back video with Ziggeo — including the handful of behaviors that are easy to get wrong and fail quietly in production.

You do not have to read the API reference first. Point your agent at the file, describe what you want, and review the result.

An Agent Skill. Claude loads it only when a request looks like Ziggeo work, so it costs you nothing the rest of the time. Works in Claude Code and the desktop app.

mkdir -p ~/.claude/skills/ziggeo
curl -o ~/.claude/skills/ziggeo/SKILL.md https://ziggeo.com/skills/claude/SKILL.md

Then just ask for what you want:

Add a Ziggeo video recorder to our application form, and store the video token against the candidate record when the upload finishes.

The same guide as an AGENTS.md — the convention OpenAI Codex and several other agents read from a repository root. Append it to your existing AGENTS.md, or paste it into a custom GPT's instructions.

curl https://ziggeo.com/skills/openai/AGENTS.md >> AGENTS.md

• The browser recorder and player — the SDK script, the ZiggeoApi.V2.Application object, and the embedding elements.
• The server API — Basic auth, and a complete three-call upload recipe that actually gets bytes into your account.
• Regions. Every account lives in one region, and calling the wrong one fails exactly like a bad key.
• Video and stream states, which are numbered on two different scales.
• Authorization tokens, including the encoding and expiry units that silently do the wrong thing.
• Transcription, and what you need before it will work at all.
• The traps: settings that are ignored when you put them on the element, the download route that returns a rendition rather than your original file, and the fields that look right and are not.

It also tells your agent what Ziggeo does not do today, so it declines instead of generating code against a feature that does not exist.

An agent acts on these instructions without checking them, so we check them for it. On every deploy, an automated harness verifies the guide against the live platform: every endpoint is probed in both regions, every SDK symbol and embedding attribute must exist in the bundle we are actually serving, and the transcription language list is compared against the transcoder itself. If a claim stops being true, the build fails.

We also test it the hard way. We hand the published file and a set of real credentials to an agent with no other knowledge of Ziggeo, and have it build a working integration from scratch. Everything it has to guess is a bug in the guide, and gets fixed.

/llms.txt indexes this documentation — the SDKs, the API reference, the use cases — as links an agent can fetch on demand, following the llms.txt convention.