anabologyco-maker/claudia-bot/taste

Taste - taps, model, tokens, then a look the owner confirms

Read docs/30-taste.md §0 (the ten lessons) and §2 first. Taste is evidence, not orders: never impose a colour thesis or a grade on the footage; ask (§4). The taste profile belongs to the user and follows them to every project.

1. Interview (§2.1): one message, the must-ask questions first (accent colour; keep the generator's colour or one grade; faces shown or hidden), then texture, density, type, taboos. Record: slop --json -p NAME decide '<their answers>' --by owner.
2. Pool (§2.4), pick one:

• their Are.na channel: slop --json -p NAME taste pool add --arena SLUG --mine (only channels they own or follow; ARENA_TOKEN only through the secure card);
• their own images: slop --json -p NAME taste pool add --dir DIR (DIR = a full path);
• nothing yet: slop --json -p NAME taste pool add --starter.

Never crawl other people's channels.

1. Embed: bash /workspace/claudia/bin/setup.sh --check --full --json must say ok. Not yet: other work, then check again (the profile's turn rule); never serve or embed before it is ok (the fallback embedder would stick to the profile). Then slop --json -p NAME job start -- taste embed --provider local (wait: the profile's turn rule). slop --json -p NAME taste status shows an embed_model that is not siglip2: slop --json -p NAME job start -- taste embed --provider local --reembed.
2. Tap app, after step 3: slop --json -p NAME taste serve --tunnel. Send the https link: "About 7 minutes on your phone: tap like or pass, pick the better of each pair, about 150 taps. Tell me when you're done."
3. Done: slop --json -p NAME taste status (taps, the meter, the stop rule: at least 150, stop at 200).

No tunnel (§2.4b): slop --json -p NAME taste sheet, send the sheets with the message it prints; the owner answers in one line; run the record command it printed, e.g. slop --json -p NAME taste record --sheet ID '1a 2b 3t 4x'.

1. Model: slop --json -p NAME taste fit, then slop --json -p NAME taste meter, then slop --json -p NAME taste tokens. Read taste/look-bible.md and ask its owner questions in one message (the evidence on both sides, a default).
2. Each answer: slop --json -p NAME taste verdict RULE confirmed --note '<their words>' (or rejected).
3. Style frames (§2.5, once plates exist): 3 frames that differ on the riskiest axes. "Rank 1-3 and say what's wrong with the others." Record: slop --json -p NAME taste verdict style_frame confirmed --value 2 --note '<their words>'.
4. Look: slop --json -p NAME taste apply (a preview), the owner's yes, then slop --json -p NAME taste apply --yes.
5. Fonts (§2.6, after a style frame): slop --json -p NAME taste fonts --tunnel, send the link, then slop --json -p NAME taste fonts rank. Only fonts the engine ships (OFL).
6. slop --json -p NAME taste stop closes the app and its tunnel.

Mood board for a shot or chapter: slop --json -p NAME taste board '<what it should feel like>'. A look complaint later ("too gray"): stop spending first, ask which frame shows the problem, record a verdict, change the look only with taste apply after their yes.