Talk Architect

Workflow

1. Intake
   • gather the source documents, venue, duration, audience, language, must-cover points, and must-avoid points
   • classify the request as single-paper, multi-paper, or research-line
2. Lock the framing
   • define the audience's core question, the talk thesis, and the main promise
   • choose pacing based on talk style and audience distance
3. Compress the story
   • reorganize around audience value instead of paper section order
   • reduce the story to setup -> development -> evidence -> resolution
4. Plan slide by slide
   • one primary message per slide
   • active English title
   • explicit transition from the previous slide
   • Chinese speaker notes by default unless the user asks otherwise
5. Keep or cut aggressively
   • keep only the strongest evidence for each claim in the main flow
   • move tables, dense ablations, and low-priority detail to backup
6. Produce handoff artifacts
   • talk_master.md for the presenter
   • slide_spec.md or equivalent structured block for downstream deck generation

Handoff

• if the story is stable and the next task is deck implementation, hand off to slidecraft
• if the visual system is underspecified, produce a short style brief and hand off to visual-language before slidecraft
• if the user later asks to diagnose an existing deck, hand off to slide-review

References

• references/style-taxonomy.md for talk types and audience distance
• references/output-formats.md for required artifact structure
• references/best-practices.md for current talk-planning heuristics and rationale