# AI interview synthesis with an agent: from recordings to themes without filling the context

> UX researchers: transcribe interview recordings on your computer and let your AI agent query 50 transcripts without loading them all. Tokenade does both.

Source: https://tokenade.net/en/solutions/data-research/ux-research · Updated: 2026-10-18

The ICP changed, so the synthesis has to be redone across fifty interviews, and loading fifty transcripts into your AI agent fills its context before the first theme appears. Add the recordings that still need transcribing and the usage limit becomes part of your research plan. Tokenade transcribes recordings on your computer and lets the agent query transcripts for the passages it needs.

Sources on this page were read on 2026-10-15; Tokenade's own figure comes from the published [THOL benchmark](https://tokenade.net/en/thol).

## What you deal with every day

Interview synthesis is one of the clearest uses of an AI agent outside software. Brittany Hobbs' [UX researcher's guide to Claude](https://productimpactpod.com/news/ux-researcher-guide-claude-tools/) (2026-04-27, updated 2026-07-27) describes Claude Code applying a coding taxonomy across forty transcripts and producing "a competent first pass in an afternoon". Four problems come with it.

### Re-synthesis is too expensive, so it doesn't happen

A product discovery practitioner who keeps [50+ transcripts in a Claude Code folder](https://elsevanderberg.substack.com/p/claude-code-for-interview-synthesis) (2026-02-04) names the old habit: "In theory, I should re-read all 50 transcripts and redraw the tree from scratch. In practice, I just start over." The agent makes re-querying possible, provided it can afford to read.

### The context window fills quietly

The [CC for Everyone research guide](https://ccforeveryone.com/guides/claude-code-for-researchers) (updated 2026-06-10) is blunt: "Do not dump 50 transcripts into one conversation. The context window fills, quality sags, and you cannot tell where." Its advice is batches of 5 to 8 transcripts per pass, then a synthesis across batches.

### Recordings are a separate chore

Many studies start as `.mp4` or `.m4a` files. Practitioners build their own transcription steps around whisper.cpp for Claude Code, as in this [published skill guide](https://github.com/Binjee/claude-code-skill-to-transcribe-audio-video-file) (2026-05-05). It works, but it is one more tool to set up per machine.

### Participant data has no built-in guardrail

Hobbs' guide is explicit that general participant PII "has no rail at all beyond your own research discipline". The same guide notes that Cowork is not covered by Anthropic's BAA in any configuration. What you feed the agent remains your responsibility.

## What Tokenade changes for interview synthesis

Tokenade sits between your AI agent and its tools, on the computer where you work.

- **Recordings become transcripts locally.** `tokenade read session-04.mp4` returns the file's details and what is said in it: from a subtitle file next to it (`.srt`, `.vtt`), an embedded subtitle track, or a local Whisper transcript if `whisper` or `whisper-cli` is installed. Transcripts are cached, so a copied file is not transcribed twice ([audio and video](https://tokenade.net/en/docs/read-documents-media)).
- **Ask transcripts instead of loading them.** `tokenade read p07.docx --prompt "onboarding friction, pricing objections, workarounds"` returns only the passages that answer, verbatim, under one heading per question.
- **Transcripts in any format.** DOCX, PDF and text transcripts arrive as clean text when the agent opens them.
- **No paying twice on a re-synthesis.** When the agent re-opens a transcript that has not changed, it gets a short pointer instead of the whole text again.
- **Nothing lost.** Long outputs are folded, and the complete text stays in a local stash, so every quote can be traced back to its source.

On long agent sessions, Tokenade is 39% cheaper than working without it, measured on the [THOL benchmark](https://tokenade.net/en/thol). On a Pro or Max plan, that is room for one more pass over the corpus instead of starting over. `tokenade gain` shows the saving on your own study.

## What stays on your computer

Transcription and compaction run on the computer where you work. Your recordings are never uploaded by Tokenade, and no transcript or note passes through Tokenade's servers; it sends usage counts only ([full list](https://tokenade.net/en/docs/privacy-telemetry)).

The transcript text the agent reads still goes to the agent's provider, with or without Tokenade. Tokenade does not replace participants' names or remove identifying details. It masks credentials it recognises; e-mail addresses and phone numbers are masked only if you set `TOKENADE_REDACT_PII=1`. Your consent forms and research ethics still decide what the agent may read.

Tokenade works with agents that run on your computer: Claude Code (including the Code tab of Claude Desktop), Codex, Cursor and others. It has no effect in Claude Cowork, and in the Claude Desktop chat only MCP results are optimised ([compatibility](https://tokenade.net/en/docs/compatibility)).

> Transcribe interviews on your computer and let your AI agent query fifty transcripts without loading them all.
>
> Install: `npm install -g @tokenade/cli && tokenade install && tokenade login` — or let your agent do it: https://tokenade.net/install.md

## Setting it up

1. Install the CLI and run `tokenade install`. It detects Claude Code, Codex, Cursor and the other supported agents ([install guide](https://tokenade.net/en/docs/install)).
2. For recordings, install `ffmpeg` and `whisper` (or `whisper-cli`). Without them, `tokenade read` still uses subtitle files exported by your call tool.
3. Open the study folder in your agent and ask for the first coding pass. Run `tokenade gain` afterwards to see what the transcripts cost.

## FAQ

### Do I get verbatim quotes or summaries?

Verbatim. `--prompt` returns the matching passages as they appear in the transcript, and the full text stays in the local stash for checking. Synthesis is the agent's work; extraction is not paraphrased.

### Is audio sent anywhere to be transcribed?

Not by Tokenade. Subtitles are read from the file or next to it, and Whisper runs on your computer. The resulting transcript is text your agent may then read, and that text goes to the agent's provider.

### Does it work in Claude Cowork?

No. Cowork runs in an isolated environment that local tools cannot reach. Use Claude Code, including the Code tab of Claude Desktop, or another supported agent on your computer.

### Does Tokenade remove participants' names?

No. It masks technical credentials, and optionally e-mail addresses and phone numbers. Names and other identifying details stay as they are, so handle them before the agent reads the files if your protocol requires it.

## Further reading

- [Claude Code subagents: what they really cost in tokens](https://tokenade.net/en/articles/claude-code-subagents-token-usage)
- [Claude usage limits: how they work](https://tokenade.net/en/articles/claude-usage-limits-explained)
- [How to reduce tokens on long agent sessions](https://tokenade.net/en/articles/reduce-tokens-long-sessions)