Imported from Nye-2/nainai-work (skills/image-drinking-chat/SKILL.md). Install upstream with npx skills add Nye-2/nainai-work --skill image-drinking-chat. Copyright stays with the author (MIT).

识图喝酒

Turn a drinking-table image into a short, realistic wine-friend group chat. The main risk is hard-guessing a bottle from weak visual evidence, so always separate image understanding from chat generation.

Required References

• Shared personas: ../nainai-skill-pack/references/personas.md
• Shared tone rules: ../nainai-skill-pack/references/anti-ai-tone.md
• Scene-card workflow: references/scene-card.md
• Quality gate: references/quality-gate.md
• Examples: examples/good.md and examples/bad.md

Workflow

1. Inspect all images or the user's image description.
2. Build an internal image_scene_card.
3. Assign confidence:
   • A: safe to name the specific wine.
   • B: only say "像/是不是/看着有点像"; do not assert.
   • C: do not mention a specific brand or series.
4. If A and current price or recent market facts are needed, search or ask for confirmed source material before writing.
5. Select 4-6 personas.
6. Generate 6-10 lines in 网名：内容 format.
7. Run references/quality-gate.md; rewrite once if any hard gate fails.

Output Contract

• Output final chat only unless the user asks for analysis.
• One line per message, format 网名：内容.
• Default length: 6-10 lines.
• At least 4 roles.
• No title, explanation, role table, image placeholder, or internal card.
• Short, uneven, casual lines.

Fact Boundaries

• Do not assert brand, series, degree, price, vintage, batch, or口感 from unclear visual evidence.
• Do not invent current prices, discounts, shop channels, gifts, or authenticity claims.
• If confidence is B or C, make the uncertainty part of the group chat naturally.