Inspiration
The script says "We doubled revenue in two years." On the recording, at 12 minutes 31 seconds, the founder says "we nearly doubled revenue in three years."
That pair comes from my test set, and it shows what an edit does to an interview. An answer gets trimmed for time. Two answers get joined. A word changes because the sentence reads better. Each change is small and reasonable, and at the end the script carries a sentence the guest never said.
Studios that make interview films and podcasts catch this by ear. Before release, someone sits with the locked script and replays the recordings, one quotation at a time.
I reached this problem late. My first design was a soundbite search, because transcription is the AI feature editors already use. An Avid survey of 120 ACE editors in 2026 put it at 60%. Then I read what ships today. Premiere searches speech inside the open project. Jumper and Peakto index a whole archive on one editor's workstation. Each of them answers "where did someone say this". I found no tool for the reverse question, "did they say what our script claims", so I rebuilt the Skill around it.
The name comes from satya, the Sanskrit word for truth and accuracy. A quotation owes both to the recording.
What it does
Setya is a Skill for UGREEN AI NAS with two jobs, and both run on one set of transcripts.
Check takes a locked script, narration or caption file and compares every quotation with the transcripts of the project's recordings. Each quotation receives one state, Matches, Differs or Not found. A clean match is counted and left alone. Every other entry is printed in full.
7. Differs. Hint: Altered. Listen at 00:12:34 in ana-interview.wav
Script: "We doubled revenue in two years."
Transcript: 00:12:31 we nearly doubled revenue in three years
The editor plays the recording at that point, marks the entry "matches the tape" or "change the script", and saves the ledger beside the script. After a revision, a re-check covers the changed lines and keeps the earlier answers.
Find answers a request such as "every time a guest talked about pricing" across all projects. It returns a selects list with the recording, the in and out times and the exact words, saved as a document and a CSV file.
The interviews stay on the NAS from the first step to the last.
How we built it
A Skill for UGREEN AI NAS is written in Markdown. Mine is one SKILL.md and four reference files. The Skill hands each step to one of three expert agents on the NAS, the file agent, the knowledge agent or the document agent, and works only with what that agent returns.
The knowledge agent transcribes each recording on the local model and saves the transcript in the recording's folder. The first line of every transcript holds the recording's path and its sha256 hash. That line is the entire index. On the next run Setya lists the folder, reads the first lines and knows which recordings are current, so each recording is transcribed once.
Two rules protect the editor from the model. Search is used only to locate a passage, and the words on screen are raw lines read from the transcript file. Every run ends with a coverage line that counts the recordings found, indexed, skipped and failed, so a partial index is visible.
Interviews are often shot as video with the sound recorded separately. Setya pairs each video with its sound file by name, or by duration when exactly one file fits within two seconds, and the editor confirms the pairs.
I checked every capability against UGREEN's public catalog. The package passes the kit's static checker with 0 errors and 0 warnings.
Challenges we ran into
My hardest problem was trust in a small local model. Comparing two texts word by word is a task such a model gets wrong, and my first version asked it for seven kinds of verdict. I cut that to three states, told the model to choose Differs whenever it is unsure, and printed both texts one under the other. A wrong state now costs the editor a few seconds of listening.
The second problem came from one parameter. When the knowledge agent saves a transcript, the NAS chooses the file name and the Skill supplies only a hint. A design that finds transcripts by name would break, so Setya recognises a transcript by its first line.
The third problem is still open. Transcription of the sound inside a video file stays unproven until I reach a real NAS, and pairing is my fallback.
Accomplishments that I'm proud of
Someone who has never seen the Skill can verify any Check result in seconds. The entry names a recording and a timestamp, and playing it settles the question.
I am also proud of what Setya refuses to say. It never calls a quotation fabricated, and it never guesses a speaker from a voice. Its wording stays at "differs from the transcript, listen at this timestamp", and the person decides.
What I learned
Writing a Skill is closer to writing a procedure for a new colleague than to programming. Each instruction to an agent has to say what to return when nothing is found, so that an empty result arrives as an empty result. Most of my revisions went into those sentences.
Naming the tools that already exist made the idea smaller and clearer. I stopped presenting search as new and put the script check first.
What's next for Spare Key
Next comes one day on a real iDX6011 Pro. My synthetic test set holds six recorded interviews, one paired video, one damaged file and a script with eight planted quotations. That day settles segment timestamps, video input, raw file reads and transcription speed.
After that I want selects in the marker formats of the main editing programs, so that a select opens at the right frame in the timeline.
Built With
- claude-code
- markdown
- ugos
- ugreen-ai-nas
- uliya

