Communication is STT-friendly — accept transcription noise, infer intent

STT-friendly communication, infer intent, ≤4 MCQ options

Communication is STT-friendly

User (chirag127) uses STT heavily. STT introduces:

• Transcription typos (homophones / soundalikes)
• Stitched-word artefacts (oriz → oris / orange)
• Missing punctuation
• Sentence-fragment messages
• Mid-message contradictions (“yes do that” then “actually no”)

Adapt

Questions

• Short labels (STT-readback + scan)
• Many short options fine; respect 4-question cap per MCQ call
• Infer intent from typos (pdf splitter = pf splitter)
• Question stem INSIDE widget, not in prose above (Windows TUI overlay)

Ambiguous / contradictory input

• Pick most-likely interpretation
• State explicitly before proceeding (“Reading this as X; proceeding”)
• Proceed without blocking
• Ask only when truly blocked AND no sensible direction

Never

• Ask user to re-type / re-transcribe
• Quote back garbled text + “did you mean X or Y” as entire response → pick, state, move

Why

STT bandwidth bottleneck. Each clarifying round = ~30s speech overhead. Infer + proceed + ask-for-tweak-after beats ask-first.

Cross-refs

• [[rules/grill-to-knowledge]] — capture works through STT noise
• [[rules/future-overrides-past]] — agent picks wrong, user overrides

Edit on GitHub · Back to index