SKILL.md
Process all transcript files in the output/ directory of the current project. Follow these steps carefully:
Step 1: Find input files
Use Glob to find all .txt files in output/. If there are no files, report that there is nothing to process and stop.
Step 2: Parse each file
For each file, extract only the content under === TRANSCRIPTION ===. Ignore the === SUMMARY === and === METADATA === sections.
Step 3: Derive the date-time header from the filename
Each filename is formatted as YYYYMMDDHHMM_SS.txt. Convert this to a human-readable date-time string including the day of the week, in the format:
## Thursday, February 26, 2026 at 2:20 PM
Use 12-hour AM/PM time. Derive the day of the week correctly from the date.
Step 4: Clean up each transcript
Apply the following rules to the transcription text:
- Correct all spelling errors
- Fix grammatical mistakes (subject-verb agreement, tense consistency, etc.)
- Add proper punctuation (periods, commas, apostrophes, contractions)
- Capitalize proper nouns and sentence beginnings
- Place each complete sentence on its own line
- Separate each sentence with a blank line
- Combine fragmented phrases into complete, coherent sentences
- Remove filler words (um, uh, like)
- Remove false starts and repeated words
Step 5: Write output
Get the current date and time by running date '+%Y%m%d%H%M_%S' in Bash. Write all cleaned transcripts to:
cleaned/transcriptsYYYYMMDDHHMMSS.md
where the timestamp comes from that Bash command. Create the cleaned/ directory if it does not exist.
The file should contain all transcripts in reverse chronological order (newest filename first), each under its ## date-time header, separated only by the headers (no horizontal rules).
Provide only the cleaned transcripts in the output file — no introductions, explanations, or additional commentary.
Step 6: Verify
After writing the output file:
- Confirm the file exists and is non-empty
- Confirm the number of
##headers in the output matches the number of input files processed
If verification fails, report the error clearly and do NOT delete any files from output/.
Step 7: Delete source files
Only if verification passes, delete all the .txt files that were processed from output/ using Bash.
Step 8: Leave the source recordings alone
Do not move, delete, or otherwise touch any .mp3 file.
The transcriber deletes each recording itself, in Python, the moment it has verified that recording's transcript is on disk.
That step used to live here, and it silently failed for months because matching a recording back to its transcript by filename is guesswork, and guesswork is the wrong thing to put in charge of deleting someone's voice memos.
Report how many files were processed and the path to the output file.