ai-exchange

name	ai-exchange
description	Use when the user asks to "check ai-exchange", "look at the ai-exchange folder/file", "fill in the noise threshold", "answer the pending video editor request", "set the threshold in the JSON", or otherwise refers to the `ai-exchange/` folder in this project. The UI writes JSON job files there; this skill fills in the `null` answer fields so the UI can consume them.

ai-exchange

The video editor UI drops JSON request files into ai-exchange/ at the project root. Each file represents one pending job waiting for your answer. Fill in the null fields according to the file's own instructions, save the file in place, and leave deletion to the UI.

Workflow

1. List ai-exchange/*.json (Glob). There is usually at most one file — the "latest" — but handle multiple if present.
2. Read each file. Every job file has this shape:

   {
     "task": "<job-name>",
     "createdAt": "<ISO timestamp>",
     "instructions": "<what to do>",
     "bounds": { ... },
     "<input fields>": ...,
     "<answer fields>": null
   }
   

3. Follow instructions verbatim. Respect bounds. Replace every null answer field with a concrete value.
4. Save with Edit — keep all other fields byte-identical. No reordering, no extra keys.
5. Do not delete the file. The UI deletes it on the next "Set from Response" click.
6. Tell the user briefly what you chose and why.

Known tasks

noise-threshold

Picks the dBFS noise threshold for silence removal.

What you get

• summary — one-sentence prose description of the audio distribution. Use this first; it's the distilled version of the stats.
• mentalModel — one of:
  • bimodal-clean — clear valley between noise and speech humps. The valley candidate usually wins.
  • bimodal-soft — two humps but the valley is shallow/broad. Quantile rules (median, median-3) usually beat valley.
  • unimodal — no clear separation; noise and speech overlap. Err toward the safer (quieter) end.
• stats — minimal percentiles: durationSec, minDb, maxDb, p10Db, p25Db, medianDb, p75Db, p90Db.
• candidates — 4–5 threshold proposals, each with simulated outcome:
  • db — the candidate threshold.
  • rule — how it was derived.
  • speechPct — rounded percent of the track kept as speech.
  • segments — how many speech segments the detector produces.
  • avgSec — average segment length.

Decision strategy

1. Read summary and mentalModel. Form a mental picture before looking at numbers.
2. Map mentalModel to a candidate family:
   • bimodal-clean → prefer valley. The structural argument is strong.
   • bimodal-soft → prefer median or median-3. Quantile rules are more robust when the valley is soft.
   • unimodal → prefer a conservative quantile (p25 or median). Don't get clever.
3. Sanity-check speechPct against content type:
   • Lecture / monologue / tutorial: 55–75%.
   • Conversational podcast with pauses: 40–55%.
   • Continuous narration / dense talk: >75%.
   • Interview with long listener segments: 30–45%.
4. Reject candidates with pathological segments / avgSec:
   • Thousands of segments with tiny avgSec → threshold is fragmenting speech. Move quieter.
   • Tiny segment count with huge avgSec → threshold is glueing noise to speech. Move louder.
5. Tiebreaker between two reasonable candidates: pick the quieter one. Too-low keeps extra silence (user can trim with "Remove Silences Longer Than"). Too-high cuts real speech (unrecoverable).

When to go off-list

Only set a custom integer outside candidates if:

• The content description contradicts mentalModel (user told you "this is mostly music with speech over the top") AND
• You can name the rule you're applying and why it beats all listed candidates.

Explain your reasoning to the user in 1–2 sentences so they can sanity-check.

Note on calibration offset

The UI applies a +10 dB calibration offset when it consumes your answer (e.g. if you write −54, the slider ends up at −44). This is a standing user preference — don't try to compensate for it. Pick the structurally correct threshold; the UI will shift it.

Extending

When a new task appears that isn't documented above, trust the file's own instructions field — it is the source of truth.