Personalise editing with exemplars
What an exemplar bank is, how importing one changes the AI editing pass, and what it costs you in privacy.
An exemplar bank is a JSON collection of real removals taken from reference edits you already made, which RoughCut injects into the AI editing prompt as worked examples of your own judgement. You import a prepared exemplars_v2.json in Settings → Models & Data, then set LLM editor exemplars above zero — the default asks for 8 when a bank is installed, and 0 turns the feature off. With no bank the editing pass simply runs plain and records that it did; the job does not fail. A bank that exists but cannot be read fails the job loudly instead of quietly ignoring it. Evaluation automatically excludes the clip under test from selection. Exemplars can carry transcript context, so treat a bank as sensitive editorial data.
On this page
Advanced, optional, and not a learning feature. RoughCut does not study your everyday jobs and does not improve on its own. An exemplar bank is a file you prepare deliberately and import deliberately, and the mining tooling that builds one is not part of the Mac app. The safe default is no bank at all, which is how RoughCut ships. To return to it, setLLM editor exemplarsto0and clickSave Configuration; the editing pass then runs with no examples, exactly as it does out of the box.
What an exemplar bank is
An exemplar is a single real removal from an edit that already happened: a short slice of transcript showing what came before a cut, what was removed, and what came after, together with the category of the removal and the language it was in.
A bank is a JSON file collecting many of those, and it exists for one purpose: when RoughCut asks the AI provider to edit a new transcript, it can include a handful of your own past removals in the prompt as worked examples. The model is being shown how you cut, rather than being told about it in the abstract.
This is prompt personalisation, not training. Nothing is learned, nothing is stored on a provider's side, and the effect lasts exactly as long as the bank is installed.
Where the examples come from
Exemplars are mined from reference edits — the same kind of manual edits that make up a golden set. They are examples of decisions a human already made and can defend.
The mining step is not in the Mac app. RoughCut can import and use a prepared exemplars_v2.json; it does not build one for you, and it does not accumulate one from the jobs you run. If you have no prepared bank, you have no bank.
The file
A bank is a JSON object with a top-level exemplars array. Each entry describes one removal:
| Field | What it holds |
|---|---|
clip_id | Which source clip the removal came from |
language | The language of that removal |
category | What kind of removal it was — a retake, a false start, filler, and so on |
flag | The mining tool's own marker for the entry |
n_words | How many words the removal covered |
before | Transcript context immediately before the removed span |
removed | The removed text itself |
after | Transcript context immediately after |
The file must be named exemplars_v2.json to be imported.
Installing one
- Open
Settings → Models & Dataand find theExemplar bank (personalization)section. - Click
Import exemplars_v2.json…and choose the file. RoughCut copies it into its own application-support storage; the original stays where it is. - Confirm the status label reads
Bank installed. - Open
Settings → Basic(orFull → Pipeline v2) and setLLM editor exemplarsto a number above0. The shipped default is8, which is what a job will request when a bank exists. - Click
Save Configuration.
Importing a second file replaces the installed bank.
How many examples are used
LLM editor exemplars is the number of examples a job asks for. 0 turns the feature off entirely; the default is 8.
Selection is deterministic rather than random, so the same job with the same bank gets the same examples:
- Only examples in the job's language are eligible.
- At most two examples come from any one source clip, so a single talkative recording cannot dominate the prompt.
- Examples are drawn across removal categories to a fixed pattern rather than all from one type, with substantial reworded retakes preferred where they exist.
More is not automatically better. Every example occupies room in the prompt that the transcript could have used, and a bank of mediocre examples teaches mediocre judgement. Start at the default.
What happens when there is no bank
Nothing breaks. With no bank installed, the AI editing pass runs plain — no examples in the prompt — and RoughCut records the fact that the bank was missing in the job's own run record. It is visible rather than silent, and the job completes normally.
This is worth stating clearly because the opposite behaviour would be a trap: raising LLM editor exemplars before you have a bank does not disable editing, and it does not fail your job.
What happens when the bank is broken
A bank that exists but cannot be used is treated as an error, not as a reason to fall back quietly. If LLM editor exemplars is above zero and the installed bank is missing, unreadable, empty, or malformed, the job fails with a message naming the requested count and the reason the bank is unusable.
That is deliberate. If you asked for personalisation, getting a generic edit that silently ignored your bank would be worse than an error — you would be comparing runs that were not what you thought they were.
The most common cause of an unreadable bank is macOS denying access to a file in a cloud-synced folder. Import from a local disk. See Troubleshooting by symptom.
Exemplars in evaluation
When an evaluation scores a clip, that clip's own exemplars are excluded from selection for the run of that clip. Without that exclusion the model would be shown the answer to the question it is being graded on, and every score would be inflated.
This happens automatically. You do not have to maintain separate banks for evaluation and for real work, but you do have to remember that an evaluation with a bank installed and an evaluation without one are different experiments. Change one thing at a time — see Read evaluation metrics correctly.
Treat a bank as sensitive
Exemplars contain transcript text: the words before a cut, the words that were cut, and the words after. That is real speech from real recordings.
- A bank mined from client work carries that client's words. Do not share it casually, and do not commit it anywhere public.
- The examples included in a job are sent to your AI provider as part of the prompt, under your own key, alongside the transcript being edited. That is the same path and the same provider as the rest of the editing pass, described in Privacy and data flow — but it means a bank's contents leave your Mac each time it is used.
- If a bank contains anything you would not send to your provider, do not install it.
Common mistakes
- Expecting RoughCut to build a bank from your jobs. It does not. The app imports and uses a bank; it does not create one.
- Raising
LLM editor exemplarswith no bank installed. Harmless, but it does nothing. Import first. - Importing a file with a different name. The import expects
exemplars_v2.json. - Mixing languages and expecting cross-language help. Only examples in the job's language are eligible.
- Setting a large number to "make it more personal". Prompt space is finite, and quality beats quantity.
- Comparing evaluation runs with and without a bank as if it were noise. It is a deliberate change to the prompt, and it should be evaluated as one.
Getting back to the default
Set LLM editor exemplars to 0 and click Save Configuration. The editing pass returns to running with no examples, and the installed bank is ignored regardless of what it contains.
Next steps
- Build a golden set — reference edits are the raw material exemplars come from.
- Bad-take removal and aggressiveness — the pass exemplars influence.
- Advanced settings reference — every configuration key, including the pipeline group.