00:00

Stack & facts — fill live; this is where guesses get corrected

Portrait — paste the client's portrait: block

Framework: one pipeline, name the stations

Adapt the stations to the client. For each: set an AUTO / ASSIST / HUMAN level + a quality gate. Everything drains into one approval queue — nothing sends itself (this is the trust layer that answers "automation will be low quality").

1
CAPTURE — standardize the input
2
EXTRACT — structured JSON, not prose
3
TRANSFORM — diffs over full rewrites
4
ROUTE — to tracker / systems
5
DRAFT MESSAGES — never auto-send
6
APPROVE & SEND — one queue, one review, human

Quality levers (reusable)

  • Diffs over rewrites — feed only the relevant section, output a patch. Kills volume limits + trivial review.
  • Golden examples — 2–3 best past outputs as few-shot anchors → consistency without more iterations.
  • Self-critique gate — model scores its own draft vs a rubric before it reaches the human.
  • One agent surface — collapse tool-juggling into a single agent that reads files & calls APIs.

Reality (fill the gaps)

Quality & trust

Appetite & scope

Decisions made

Action items (owner = who)

Built / demoed live

Show-don't-tell cues — energy shifts when they SEE it work

0 / 5

Openers if they stall

  • "Show me the last thing you did by hand — let's automate that one live."
  • "If only one piece got solved today, which?"
  • "What would make you trust the output enough to send after a 30-second glance?"

Live notes

Follow-ups I owe / they owe