expectedwrong hindsight

The Transcript Trick

Feeding Claude Code your own words turns out to be the most obvious thing nobody told you to do.

2 min read 361 words #claude-code #workflow #ai #productivity
hindsight — nailed it

Feeding transcripts to Claude Code changed the entire conversation. The amnesiac prompting-from-memory approach was always the wrong way. The transcript trick is now standard practice.

I started giving Claude Code my meeting transcripts last week and I don't know what I was doing before that.

The thing is — I spend hours in calls where we figure out what to build, why we're building it, what the edge cases are, what we tried that didn't work, what the customer actually said versus what we thought they said. All of that lives in a transcript that I then... don't use. I walk away with a vague sense of decisions made, open a editor, and start prompting from memory like some kind of amnesiac.

Feeding the transcript in changes the entire conversation. CC stops being a code-completion tool and starts being something closer to a collaborator who was actually in the room. It knows the constraint that came up at minute 34. It knows why we ruled out the obvious solution. It doesn't need me to re-explain any of it.

The prompt is basically nothing. "Here's the transcript from this morning's call. Let's implement what we decided." That's it. That's the whole thing.

What I didn't expect is how much it improves my own thinking about what we decided. Half the time I'll paste in a transcript and CC will surface an ambiguity — two people said contradictory things twenty minutes apart and nobody caught it live. The transcript holds the receipts in a way my memory absolutely does not.

I'm now doing this with user interviews too. You talk to a customer for an hour, you get a transcript, you paste it into the context before any feature work that touches their use case. The feature comes out different. Better. Less "what we assumed they wanted" and more "what they actually said, three times, in slightly different ways."

The obvious objection is that transcripts are long and context windows are expensive. Sure. But they're also dense with exactly the kind of ambient knowledge that makes the difference between code that technically works and code that was built for the right reasons.

The less obvious objection is that you have to actually record things. Which means admitting that your memory is bad. Which is, it turns out, fine.