tech-talk playbook / ai-literacy

Give an Engineering Tech Talk
like a Staff Developer-Advocate

A worked rundown for engineers who want internal deep-dives to land — rehearsed against Open Design's own agent-stream, sandbox, and artifact internals.

Platform DevEx Guildinternal eng talk · 45 min + Q&A
#devrel #eng-talks #agent-stream

agenda.toml

Today's roadmap

01Why now: internal talks are the highest-leverage devrel channel~5min
02Workflow map: the 6-beat rundown that keeps the room with you~10min
03Pilot: rehearsing the talk on the agent-stream internals~12min
04Risk control: the demo-fails-live playbook~8min
05ROI + rollout: from one talk to a recurring series~5min

// why-now

The problem: engineers stop trusting
demo-slides, not demo-code.

Slide decks

Static bullets compress a whole system into buzzwords; senior engineers tune out by slide three because there's nothing they can falsify.

❌ nothing to verify

Marketing-style demos

A polished screen recording proves the feature ships, not that you understand why it works — the gap shows up the moment Q&A starts.

😩 no mental-model transfer

Staff-level walkthroughs

Show the real code path live — the SSE stream, the sandboxed iframe, the artifact write — and let the room poke at it. Trust compounds instead of decaying.

✅ engineers verify, then believe

workflow map · 1 / 2

Map the talk before you write a single slide.

Pick one running system, then plot six beats against it: why now, architecture, live walkthrough, failure modes, impact numbers, next steps. Every beat should end on a question the audience can actually ask out loud.

6 beats 45 minutes one live system
rundown.md
# engineering tech-talk rundown
$ talk map --subject agent-stream

  ✓ Why now         ✓ Architecture map
  ✓ Live walkthrough ✓ Failure modes
  ✓ Impact numbers   ✓ Next steps
  …

// 6 beats, 45 minutes, one running system
// every beat ends on a question the room can ask live

pilot · 2 / 2

Pilot the talk on a system you actually run.

The pilot subject is Open Design's own claude-stream.ts SSE loop, the sandboxed iframe preview, and the artifact write path — nothing hypothetical, nothing slideware. Three colleagues sat in on the dry-runs before the real room.

✦ 3 dry-runs before the real room
✦ timed to the second, per beat
✦ live terminal, no recorded video
✦ audience holds questions to the end

pilot · 5 checkpoints

1. Pick the system — the agent-stream SSE loop, not a hypothetical one
2. Script the failure — kill the daemon mid-turn, narrate the reconnect live
3. Time-box each beat — 45 min hard stop, timer visible to the speaker only
4. Run it past a skeptic — one senior engineer allowed to interrupt
5. Cut anything that needs a slide to explain what the code already shows

talks/agent-stream-deep-dive/risk.md

Every live demo needs a written recovery path.

risk.md
## Known failure modes
- daemon restarts mid-stream:  12s reconnect — narrate it, don't hide it
- sandbox iframe goes blank:   fall back to the 90s recorded clip
- Q&A runs long:              cut the "further reading" slide, keep the room

## Rule
- Never debug live past 60s; switch to the fallback clip, explain after

# rehearsed 3x against this exact list before the real room
# the backup clip is muted b-roll, not a scripted answer

// roi

Three numbers that justify the rehearsal time.

1 · Time to competence

23 engineers who watched the sandbox walkthrough shipped their first agent-stream patch in under a week, down from roughly three.

2 · Fewer repeat questions

"How does the SSE reconnect work" pings dropped from ~9/week to ~2/week once the recording shipped to the wiki.

3 · A reusable format

The same 6-beat rundown ran twice more this quarter — sandbox isolation, artifact storage — each under 4 hours of prep.

Recording + rundown: internal-wiki/agent-stream-talk · #eng-devrel

>_

Turn one good talk into
a recurring rundown

One system · one rehearsal · one recording engineers actually rewatch

next: sandbox-isolation talk · Q3 wk2 owner: platform-devex guild