Hand a Jam to your agent - Jam Documenation

Today you describe a bug in a sentence, maybe with a screenshot, and the agent guesses which page you mean. Paste a Jam link instead. The agent gets what the person said while it happened, the request that failed with its status and body, the click before it, and the frame at that second. The same works when the Jam shows a feature someone wants, a flow they walked through, or a question. You get back what fits: an explanation, a reply, a fix, or every change they asked for. Each points to a moment in the recording that you can check.

To explain or draft a reply: any client with Jam MCP connected, such as Claude or ChatGPT.
To fix or build: a coding agent with your code open, such as Claude Code, Cursor, or Codex, and the Jam skills. Install the Jam plugin, which also connects Jam MCP (see Connect your client), or run npx skills add jamdotdev/skills.
To act on a walkthrough: the Jam CLI as well, signed in.

Steps

Paste the link and say what you want

Put a few words before the link. They decide what you get back:

An explanation. Say “What is this person running into?” You get the problem in plain language, with the moment it happened.
A reply. Say “Draft a reply to them.” You get two sentences to paste into the ticket or the chat.
A fix. Say “Fix this.” You get a patch aimed at the cause of the failure.
Every change from a walkthrough. Say “Do what I said in this Jam.” You get a numbered list of the asks, then the changes.
The agent’s pick. Paste the link alone in a coding agent. A bug gets a diagnosis, a walkthrough gets a task list, and an unclear Jam gets one question back.

Check the answer against the Jam

A good answer points at a moment: a quote from the person, a timestamp, or a line like “POST /api/checkout answered 500 at 42 seconds”. Open the Jam at that second and confirm you see the same thing. An answer that names no moment is a guess. Ask the agent where in the Jam it saw it.For a walkthrough, the agent posts the list first, one block per ask, in the order the person said them:

T3 · 1:42 · visual
Said:  "the price on this card is cut off"
Frame: frame-102000.jpg, the Pro card's price row
Plan:  let the price wrap instead of truncating

For any ask with two readings, it asks you and starts on the rest without waiting. Answer its questions. Then check that every ask you remember is on the list, and that each frame shows what the person pointed at.

Review the fix or the feature

A fix changes the cause. A swallowed error or a retry around the failing request is not a fix, so send it back. A feature should match what the person showed, screen by screen. A walkthrough ends with a count, such as “11 asks: 9 done, 1 deferred, 1 needs an answer”, and every ask that is not done says why. With the Jam CLI installed, say “Record a Jam showing the fix works.” The agent records the change running and returns a second Jam link as proof. See Prove a fix.

Check what the agent left on the Jam

Open the Jam. If the agent left no comment, say “Comment on the Jam with what you found and mention the person who recorded it.” The comment appears under your name, followed by “via” and the agent’s name, such as “via Claude Code”. The mention notifies the person who recorded it when they are in your workspace.When there is a pull request, the Jam’s Info lists it under Resources. If it does not, say “Attach the pull request to the Jam.” The next person who opens the Jam goes straight to the fix.

FAQ

What does the agent read?

The transcript, when the person spoke. It is the report, and it often names the problem before the logs do. The agent reads the console errors and the failed requests, with their status, body, and the second they happened. It also pulls frames from the video around the failure, which show what the person saw: the card field filled, the button grayed out. For a walkthrough, it reads the whole transcript and pulls a frame at every “this” and “here”, because the cursor shows what the person pointed at.The details also carry the Jam’s links: the tickets and pull requests attached to it, and the Jams that fix it, prove it, or relate to it. So the agent sees a fix someone already linked before it starts its own.With Jam MCP these are tools: getDetails, getVideoTranscript, getUserEvents, getConsoleLogs, getNetworkRequests, and getFrames. With the Jam CLI they are jam get commands, the comment is jam create comment, and the link is jam update jam --resource:

jam get jam abc123
jam get network abc123 --status 5xx
jam get console abc123 --level error
jam get frames abc123 --at 42000
jam create comment abc123 "500 on /api/checkout: totals missing when a coupon is applied. Fix in #1423." --at 42000
jam update jam abc123 --resource https://github.com/acme/app/pull/1423

Every filter and frame option is in Read Jam data.

What does the agent do with a bare link?

With the Jam skills installed, the jam skill reads the Jam’s details first, and the transcript when the person spoke. Several asks, whether in the narration, title, description, or comments, become a task list. One bug, an error, or a failed request gets a diagnosis, with no code changed. When it cannot tell which applies, it asks one question, such as “Fix it, write a test for it, or file a ticket?”

The Jam has no voice or no video

A Jam recorded with the microphone off has no transcript, so the console and network are the report. A screenshot Jam from the browser gives the agent the image, the console, and the network. An Instant Replay has no frames or screenshot, so the agent works from the clicks, the console, and the network. When you say “Do what I said in this Jam”, the agent takes the asks from the title, the description, and the comments.

The agent went to the logs and ignored what the person said

Say “Read the transcript before the logs.” Ask for one entry per issue: the quote, the frame that shows it, and the log lines from the same moment. One narrated Jam often carries several issues.

The agent says the transcript or the frames are not ready

They take a few minutes to process after upload. The agent works from what is ready and says what is missing. Ask again in a few minutes for the rest.

I have a folder of Jams, not one

Name the folder in your request, such as “Fix the Jams in the Checkout folder.” The agent finds them by folder and reads each one. Ask for one fix per cause, so several reports of the same bug get one fix.When the folder holds many unrelated problems, sort them first. Turn Jams into tickets gives you one ticket per cause, and each ticket comes back here.

​Steps

​FAQ

Steps

FAQ