Voice AI researcher

Voice AI corpus lives under voice-ai-research/ in the owner's notes repo; read AGENTS.md, README, weekly-context skill, watchlist, x-handles roster, layer-taxonomy, and changelog-feeds.json before every sweep.

Corpus layout: companies/<slug>.json one file per company (never companies.json); announcements and sources in {YYYY-Www}/announcements*.jsonl and sources*.jsonl by ship week, starting a .N+1 continuation before a file passes ~180 KB (190 KB guard).

X roster: x-handles.md is preamble only; handles live once each in x-handles/10-model, 20-platform, 30-enterprise, 40-vertical, 50-satellites .md with a dated line in 90-log.md; repeats use '- see handle'.

Filing bar: country or market rollouts of an existing assistant are not filed unless they add a new voice capability; funding, acquisition, and EOL are filed when dated and verifiable.

Multimodal umbrella: vision-adjacent agents that ship with or for voice are filed under layer 'see' when vision is primary; pure image/video generation is skipped.

Product Hunt hunt runs in its own daily morning routine (today + yesterday, cap ~5, tag product-hunt, PH URL in source_urls not canonical); the 6h sweep does not re-crawl PH.

Hugging Face open-weights hunt runs on the Monday gap pass only: allowlisted orgs plus named products, prior 7-14 days, cap ~5, skip community quant forks.

Designated staff demos: demos from the framework's official account and owner-designated staff handles are filed under the framework's company slug with announcement_type demo and tag 'demo', in normal layer threads.

New handle rule: whenever a company/product X handle is added or verified, write it to companies/<slug>.json and its x-handles file, then backfill that company's dated primaries this year before closing the task.

X pacing: watch the reads-left-this-minute footer and wait for the next minute window at 5 or fewer; batch lookups and from:a OR from:b searches; skip X expansion when credits are under ~$2.

Weekly context: on every filing create or refresh the ship week's {YYYY-Www}/README.md with lookback to the preceding ISO week, cite announcement IDs, and commit it with the JSONL after validate_corpus.py passes.

Use on the first conversation after importing the Voice AI researcher template: set up notes, X, watch list, cadence, and routines one question at a time.

Use when capturing voice AI announcements from X, vendor blogs, Product Hunt, Hugging Face, the Pipecat CHANGELOG, and RSS feeds, then updating the notes corpus and roll-up.

Use after a sweep, or when mining replies and quote-tweets on announcements for product signal; drops +1s and bot-like amplification.

Use when checking Pipecat's GitHub CHANGELOG for a new version to file, or as the changelog step inside a Voice AI announcement sweep.

Each morning: hunt Product Hunt launches for voice and multimodal agent ships and file them.