Don't dump the meeting transcript into the model

A kickoff meeting runs an hour and the transcript comes out tens of pages long — side conversations, half-decisions, requirements buried inside debates about something else. The obvious move is to hand the whole thing to a model and ask for the brief, the action list, the first cut of requirements. We tested this on a real transcript before trusting it on client work, and the failure wasn’t where we’d been looking.

Recording decisions is the easy part

Asked to summarise what a meeting decided — what was agreed, what was reversed, what got dropped — a modern model is reliable. Reversals tracked correctly. Corrections honoured. Dead features filed as out of scope. For the narrow task of recording what happened, this is solved. It’s also not where the cost is.

The cost shows up when you ask it to do the work

The moment you ask the model to do somethingwith the transcript — draft the auth requirements, scope the dashboard, write the data model — the whole transcript is sitting in its context. Most of what’s in there is irrelevant to the task in front of it. The dashboard tangents aren’t neutral filler when you’re working on auth; they’re distractors competing for attention. The integrations debate isn’t background when the task is the data model; it’s noise the model has to reason past.

This isn’t speculation. There’s a steady research thread — context rot, lost-in-the-middle, distractor density — showing that model performance degrades as irrelevant material fills the window. The signal-to-noise ratio of the context matters as much as its size. A fifty-page transcript with two pages of auth-relevant material is not a richer context for an auth task than the two pages would be on their own. It’s a worse one.

The fix is structure, not a longer window

The answer isn’t a bigger model or a larger context window. It’s the discipline every serious agentic-coding setup already lives by: keep the context tight to the task. Pre-process the transcript into task-aligned views — decisions, action list, requirements grouped by area, open questions, named owners. When the task is auth, only the auth slice goes into the model’s window. When it’s the dashboard, only the dashboard slice does.

The work is in the structuring. Deciding what the right slices are, what counts as a requirement versus a constraint, who actually owns what — these are judgments that don’t survive being skipped. Dumping the raw transcript into a model looks like it’s saving that work. It’s loading it onto the model at the worst possible moment, while the model is also trying to do the downstream task.

The transcript isn’t the artefact

The shape of the failure isn’t that AI can’t read a long document. It’s that giving a model a long, unstructured, off-task source and asking it to do precise work degrades the work in ways that don’t always show up as obvious errors. The kickoff transcript isn’t the artefact you want in context — it’s the raw material for the artefacts you want in context, and producing those is most of the job.