Why this skill exists

It started with a calendar full of recurring client calls and a pattern nobody enjoyed. Every call ended the same way: someone owed notes. Writing them took half an hour the meeting had already spent once. Skipping them meant decisions lived in memory, and memory is where agreed deadlines go to die.

Zoom was already recording the transcripts. AI Companion generates a full word-for-word record with speaker labels, and it was sitting in the account unused. The gap was never the technology. The gap was the plumbing: fetch the transcript reliably, turn it into notes a person wants to read, and file those notes where the team keeps things. Every week. Without anyone remembering to do it.

So we built the plumbing once, hardened it against everything the Zoom API threw at it, and turned the whole thing into a skill any agent can follow. Then we released it free, so your agent can do the same thing for your calendar.

What the skill actually is

A SKILL.md your agent reads. It documents the two separate Zoom transcript systems, the exact fetch chain, and eleven pitfalls we hit so your agent does not have to. A stdlib-only Python script that handles OAuth, retries, and dedup. Four note-style templates that turn raw transcripts into notes with verbatim quotes and no invented deadlines. An onboarding guide that gets you from zero to a scheduled fetch in about fifteen minutes.

What the Zoom API taught us (so your agent doesn't learn the hard way)

• Zoom has two transcript systems, and querying the wrong one silently returns nothing. Cloud Recording transcripts and AI Companion transcripts live behind different endpoints. Accounts that record locally often have an empty recordings endpoint by design. The skill checks which system the account uses before debugging anything.
• The scheduled meeting UUID is the wrong UUID. Asking for a transcript with it returns error 3323, "transcript does not exist," even when the transcript is right there in the portal. The working path is the instance UUID from the instances endpoint. This was the single most common failure during the build.
• Some UUIDs need double URL-encoding. An instance UUID containing a slash 404s unless it is encoded twice. The script handles it. A hand-rolled request usually does not.
• The instances response key is "meetings," not "instances." A parser looking for the obvious key finds nothing and reports zero meetings. This one cost an afternoon.
• A meeting does not exist for the API until it ends. A scheduler that fires while the call runs long finds nothing and exits. The script keeps polling inside its wait window instead, and retries for up to 40 minutes while Zoom processes.
• Dedup is what makes scheduling safe. Every fetched transcript is recorded in a state file, so each one downloads exactly once and a missed run is picked up by the next. The script is silent when there is nothing new and loud only on real errors, which is what a scheduled job should be.

Why it's free

Because the hard part was never the model. It was the workflow: the right endpoints in the right order, the retry discipline, the note formats with real fidelity rules. Most owners have an agent that could do all of this. They just do not have the tested sequence and the guardrails. Now they do.

Who maintains it

The skill is maintained by My AI Freedom Systems, a small team that builds practical AI workflows for real businesses and releases the reusable parts in the open. Everything published under that name is scrubbed of client data and tested against real work before it ships.

Links

• zoom-transcript-fetch · this skill, free on GitHub
• harvest-testimonials · find every real client testimonial you have already earned
• graphic-testimonials · turn those testimonials into branded graphic cards
• freedomaios.com · the parent site