Write an editorial skill that finds interview selects
Turn a working producer's checklist into a reusable SKILL.md with source timecodes, honest gaps, and a small test interview.
A producer asks for three interview selects for a film about a neighbourhood library. A language model returns three elegant sentences. One belongs to the interviewer, one combines halves of two answers, and one has no source time. All three look fine in a document. None belongs on an editor's timeline in that form. The problem is visible as soon as someone tries to use the output.
A reusable editorial skill can help if you first define what a usable select looks like. This example concerns one tightly bounded job. Someone supplies a timecoded transcript and a short brief. The skill identifies candidate passages for a human editor to audition against the recording. You can try the whole design on a few paragraphs of invented dialogue before loading any real interview.
Begin with the editor's next action
An editor opening the selects list should be able to find the source clip, play the passage, and decide whether it advances the story. That suggests the necessary fields. Include the clip identifier, speaker, start and end from the supplied transcript, exact excerpt, and one sentence explaining why it might work. If the timecode marks are only paragraph-level, identify the containing range. Exact-looking frame values invented from a paragraph timestamp create extra work at the next step.
Give the exercise a brief. The library film follows a teenager repairing donated books after school. The desired opening introduces the damaged object, the middle explains why the repair matters, and the ending stays with an observed action. The brief need not contain a final script. It gives the selector a reason to choose one precise answer about repairing a spine instead of a broad answer about loving books.
Turn one good human select into an example
Write a tiny sample transcript. In clip LIB-03, from 00:10 to 00:18, the interviewer asks what first caught the student's attention. From 00:18 to 00:26 the student replies, “The cover was hanging by one thread. I could see the old blue glue.” From 00:26 to 00:34 the librarian says that some books arrive in better condition than others. These are invented lines for a test.
A useful opening candidate names LIB-03, the student, and the 00:18 to 00:26 interval. It quotes the student's words exactly and explains that the detail of blue glue gives the viewer an object to look for. The final explanation is editorial judgement and belongs in its own field. It should never slip inside the quotation marks. Include this input and output together. A concrete pair resolves several ambiguities that the phrase “choose vivid quotes” leaves open.
Give the skill a clear boundary
The Agent Skills format puts the instructions in a file named SKILL.md inside a directory named for the skill. The required metadata includes a name and description, followed by Markdown instructions. For this example, interview-selects is a workable directory and skill name. The description could say that it proposes traceable interview selects when a user provides a transcript and a story brief. Someone asking for a release form or sound mix is asking for a different job. Agent Skills specification
In the body, describe the sequence in ordinary language. Read the brief, identify which speakers can supply evidence, locate contiguous passages, and rank candidates by their value to the proposed story. Specify what to do with weak material. If the transcript contains no ending that meets the brief, return the strongest available middle and identify the missing action. A heading left honestly incomplete is more useful than a quotation attributed to someone who never said it.
Give hard cases a predictable outcome
Consider an answer that starts with “that was the day it changed.” Read the preceding exchange. If the word “that” needs an earlier sentence, widen the candidate to include the referent or flag that it requires context. If a speaker repeats a claim that the interview cannot verify, identify it as the speaker's claim. A good select may still need corroboration before it becomes narration or an on-screen fact.
Plan for imperfect transcripts too. When two speakers are both labelled “Speaker 1,” preserve the uncertainty. When there is a gap in the timecodes, return the words and clip identifier with a note asking the editor to locate the boundary by ear. If an answer includes names or private details irrelevant to the story, choose another contiguous passage where possible. For testing, use invented or appropriately cleared material that you can safely hand to the chosen service.
Test the output as if you had a timeline open
Run the short library example, then change one thing at a time. Remove the timestamps. Swap the labels on the question and answer. Add two similar quotes in different source clips. Ask for four strong candidates when the transcript supports only two. For each result, check the exact words, the speaker, and whether the supplied locator takes you to the correct passage. Keep a short list of defects you can reproduce.
If one test fails, rewrite the smallest relevant instruction and rerun the other cases. Suppose the tool keeps combining sentences across an interruption. Add a concrete example showing where a contiguous passage ends, then check that ordinary uninterrupted answers still survive intact. The official best practices for skill authors also encourage developing instructions from real work and checking how they behave in use. Agent Skills authoring guide
Let the editor listen before approving
An accurate transcript cannot tell you everything about an answer. A laugh may be nervous, a pause may carry the real thought, or an apparent statement may be a question in the speaker's delivery. Put the candidates into a temporary sequence with handles on both sides and listen. Record which selections survived and why. That feedback can improve the next revision of the skill and the questions asked on the next shoot.
Keep the main instructions short enough to read in one sitting. A single checked example and a clearly described failure case will usually earn more space in the file than another paragraph asking the system to be excellent. The quality test remains pleasantly concrete. Can an editor find what was said, hear it in context, and choose what belongs in the film?
Sources
- Required metadata, file structure, and references. Agent Skills specification
- Extracting expertise from a task and evaluating the result. Agent Skills authoring guide