Talk track
- 01Opening
It’s 2026 and coding looks a lot more like this.
Use the actual intro from the article. More and more, engineering is talking to machines. Let the prompt demo fail; that sets up the rest of the talk.
- 02The question
What does it mean to be good at talking to Claude, Codex, or whatever comes next?
This is the question at the end of the intro chapter. The structure should feel like the article: mindset, equation, techniques, map, orchestration, recap.
- 03The plan
Chess pawn
Let the structure become the table of contents for the talk.
- 04Section 01
The beginner’s mindset.
Section break. This starts the mindset chapter: we are early, and the people still learning are going to pull away.
- 05Chess is 500 years old
AI coding is barely 3 years old.
Use the pawn to make the chess analogy concrete. The point: the meta has barely been uncovered.
- 06Beginner mindset
Three years in is about 1200 ELO.
Use the chess analogy from the article. The important bit is: do not trust the feeling of competence too early.
- 07Two responses
Sooner or later the model will let you down.
This slide should be mostly visual. Closed: blame the AI. Open: ask what you could have done differently. That is the article’s core posture.
- 08Gaming culture has a useful phrase
Skill issue.
The article’s strongest mindset line. When your character keeps dying, the level isn’t broken — you’re missing something. Sounds flippant; the move underneath is serious.
- 09Practice curve
Every useful failure can become a little ELO.
Like doing one chess puzzle every day. Try something new, notice what worked, update your repertoire. Or: close the question with “the model is bad,” stop practicing, regress.
- 10Pragmatic optimism
“I suggest just being 2 or 10 times more ambitious than before, because it might just work.”
Pragmatic optimism strategy. The point is not blind hype. It is a useful prior: try the slightly more ambitious thing before assuming the boundary is fixed.
- 11Q
When the work disappoints you, what do you usually do?
Open-ended — let people share whatever shape this takes for them. Listen for whether they look outward (blame the thing) or inward (what did I miss).
- 12Section 02
Pull the levers.
Section break. Shift from posture to diagnosis: what part of the setup produced this output?
- 13The simplest frame is an equation
Better inputs. Better outputs.
Use the article equation. First AGENT × INPUT, then expand it. This is the frame everything else hangs on. Walk the four levers when output disappoints.
- 14Tool
Tools are changing every day.
Most of us are already on agent-native tools. The point is: don’t bolt AI onto VSCode. Cursor, Claude Code, Codex slice context smarter, manage long tasks, persist state, and inject project-aware system prompts. Tool is the smallest of the four levers, but it’s the easiest one to pull.
- 15Model
Use the best model.
Pick the smartest model the work warrants. Walk down the article guide: fast/Sonnet, taste/Opus, hard/GPT-5.5, ambiguous/thinking on.
- 16Effort
The cheap dial nobody touches.
Thinking effort is becoming one of the most important levers. Low for trivial edits, medium for executing, high for real decisions, xhigh for novel problems. And: stop using Auto mode. It’s optimized for platform margin, not your output.
- 17Prompt
Clarity is everything.
Walk the article comparison: “make this faster” vs “first paint is 2.4s, target under 1s, profile and start with the biggest wins.” Then the don’ts: “be careful,” “think step by step,” politeness padding.
- 18Context
When in doubt, load more context.
Load: project rules, related PR, design doc, screenshots, the failing trace, MCPs into your actual systems. If you’re rewriting the same prompt for the fifth time, stop — the lever you need is the next one.
- 19Persistent context
Don’t repeat yourself to the agent.
Project rules are the longest-leverage context you can write. CLAUDE.md, AGENTS.md, .cursorrules — the same paragraph that gets a new hire from “lost” to “useful” in a week does the same for the agent. Capture the things you find yourself correcting twice.
- 20Q
When did you realize you had a lever to pull, and outputs got suddenly better?
Open-ended — the "aha" moment. Look for the specific lever they discovered: switching tools, turning thinking on, dropping in a real doc, writing a CLAUDE.md.
- 21Section 03
Collect moves.
Section break. This is the techniques portion of the article: prompting as a repertoire of small moves.
- 22Techniques
Build a repertoire.
This comes straight from the techniques chapter: openings, tactics, pressure-test, pay it forward. Keep the slide light: the moves are grouped so the structure is part of the lesson.
- 23My new favorite
YOLO and extract.
My new favorite technique. Have agents complete a huge chunk of work in a single PR — a whole milestone, not a single change. Then treat the PR as disposable: extract the parts you like as smaller PRs, or as insights. The dial is hands-on (ask along the way, edit the plan) versus YOLO (let it decide). Spec-heavy work goes hands-on; open-ended exploration goes YOLO. Load lots of context — docs, Figma, prior PRs, RFCs. Tell it to commit incrementally so you can cherry-pick. GPT for the run; Opus for the PR description. Trust that something in there is useful, even when most of it is wrong.
- 24Q
What techniques do you keep coming back to?
Open-ended — patterns people lean on, not just prompts. Could be openings, recovery moves, things they stole from a teammate, things they've done since forever.
- 25Section 04
Explore the map.
Section break. The next step after collecting moves is knowing where each move belongs.
- 26The map
Every change lives somewhere on a 2D map.
Use the interactive article map. The axes are breadth (which area of code) and depth (how zoomed in). At any cell, three moves: ask, plan, or delegate. Click around so the room sees the prompt change with the cell.
- 27Choosing your move
Pick the move that fits the cell.
Tie the map to the three moves with real examples. Ask when you don’t know yet. Plan when you know what but not how. Delegate only when the path is clear and the result is gradeable. Calibration is most of the skill.
- 28Q
What’s your mental model for talking to the AI?
Open it up. Listen for metaphors people already use: intern, pair programmer, search engine, compiler, reviewer, weird teammate.
- 29Section 05
Conduct the fleet.
Section break. This starts orchestration: from one prompt at a time to managing several runs.
- 30Orchestration
Think in hours, not messages.
Use the article’s pinging-versus-full-prompt example. The model loses nothing by knowing the destination; you lose four context switches.
- 31Chain, don’t ping
Same work. Less idle time and context switching.
This is the first orchestration visual. The grey blocks are you deciding what to type next. That is the hidden cost.
- 32Stagger and overlap
They run together. You can’t.
Use the article’s cascade idea. Purple is you and never overlaps. Pink is the agent and overlaps freely. Grey is the small but real cost of switching attention between threads.
- 33Fan out
Try spawning sub-agents.
Good candidates: tests beside implementation, unrelated bugs, library spikes. Bad candidates: chains where B needs A’s output. Budget for the merge — three branches become one review queue.
- 34Long jobs
Long runs need a contract, not babysitting.
Talk about the instinct to hover. If the task can survive without you, the agent can return a branch instead of a half-finished conversation. The contract is goal, clarity, success criterion, judgment boundary, and fallback.
- 35Q
How do you manage many agents at once?
This should pull out practical habits: naming threads, staggered task descriptions, review queues, keeping scope boundaries sharp.
- 36Section 06
Keep climbing.
Section break. This is the recap: we are early, the skill is real, and there is a lot left to learn.
- 37Six reminders before you go
The whole talk in one page.
Six recap cards — one per chapter — each with the chapter’s single sharpest line.
- 38The next 1200 is wide open
Keep climbing.
Final card. Close with a direct invitation to practice.
- 39Q&A
Q&A
Open floor. Seed prompts if it goes quiet: what surprised you, what would you push back on, what would you ship tomorrow.