Reference · print this
Speaker's card for the 19 slide deck. Audience: professional developers, almost no agent experience. The deck teaches the basics of agentic development, not the skills catalogue.
| Min | Slides | Beat | Say this |
|---|---|---|---|
| 0–2 | 1–2 | Disarm the room | This is not a machine learning topic. Nobody has to train anything. Their engineering judgement is the asset and it transfers. Say it plainly; it decides whether the next 23 minutes land. |
| 2–6 | 3–4 | What an agent is | The foundation slide. Slow down here. A model, a set of tools, a loop. Then the one distinction: in a workflow the code decides when to stop, in an agent the LLM decides. The deck no longer says "less predictable"; say it yourself. The agent trades predictability for flexibility, and that trade is why constraints matter. |
| 6–9 | 5–6 | No memory | "An agent can only work with what's in its context window." Then the frame they keep: "AI is not a super-powered developer. It's a new starter with no memory." Everything after this is briefing a new starter. |
| 9–15 | 7–10 | Habit 1 · plan mode | What the mode does and does not allow. Why it works twice: it primes the agent's context, and it shows you what you actually want. End on the line about vague requirements, and pause. |
| 15–20 | 11–14 | Habit 2 · the loops, then hooks | "AI agents don't get frustrated by repetition." A red test is an instruction, not a complaint. Then the four loops; ask the room which one their repo is missing. Say the difference aloud before slide 14. The "Hooks" loop on slide 13 is a git pre-commit hook: it fires on a git action and blocks the commit. Slide 14 is a Claude Code hook: it fires on the agent's own lifecycle. If you do not say this, the room will mix the two up. End on the beginner pair, PostToolUse with an Edit|Write matcher, and Stop. |
| 20–22 | 15–16 | Stay the engineer: the codebase, then skills | The codebase matters more than the prompt. Your existing design skill is what makes an agent useful here. Then skills: a skill is only the process written down, and three sentences are enough. Move fast; these two answer a question the room has not asked yet. |
| 22–25 | 17–19 | Close: nothing new, Monday, end | None of this is new. It is twenty years of good design, pointed at a new kind of colleague. Then Monday: one ticket in plan mode, one feedback loop, read the diff every time. |
Cut slide 16, skills, first. Cut slide 15, the codebase, second. Both answer a question the room has not asked yet. Never cut slide 3 or slide 4, what an agent is. Never cut slide 10, the vague-requirements line. Never cut slide 14: it is the only slide that tells the room what to actually install.
Run plan mode live against a real ticket in front of the room. Let them watch it ask clarifying questions. That is the most convincing thing in the session. Second choice: open a real repository and show what a blocked pre-commit hook sends back to the agent.
| Question | Answer |
|---|---|
| "Does this replace me?" | Go back to slide 2. The agent has no memory and no opinion about what is worth building. It also cannot tell you the plan is wrong. |
| "Why not just give it the whole task?" | Slide 4. The LLM decides when to stop. More scope, less predictability. Constraints are the fix, not a longer prompt. |
| "It keeps forgetting what I told it." | Slide 5. Fresh context window. It never knew. Put it in the plan, or in a file it reads. |
| "We are not a TypeScript shop." | The four loops are examples, not the point. Any automatic pass-or-fail signal works: a compiler, a linter, a test command, a hook. |
| "Where do I start on Monday?" | One real ticket, plan mode, no code until the plan is right. |
| "How is a hook different from a git pre-commit hook?" | Slides 13 and 14. A git hook fires on a git action, so it can only block the commit. A Claude Code hook fires on the agent's own lifecycle, at PostToolUse or Stop. Only the second one gives the error back to the agent while it is still working, so the agent fixes the problem itself. |
Accuracy note. Every quotation in this deck was fetched
from the Markdown twin of its page on 2026-08-27 and checked character by
character. The record is in RESEARCH.md and
CONTENT-BRIEF.md. If someone challenges a quote, the source is on
the slide.
One quote from the earlier deck was removed: the "fleet of middling to good engineers" line, cited to the 7-phases page. That text is not on that page. Do not use it.
The instruction-budget slide is removed. The hooks slide took its place on 2026-08-27. The ~150-200 instruction figure is therefore not on any slide, and no slide credits it. Keep the figure as background for a question from the room only. If you use it, attribute it correctly: it is Kyle from Humanlayer, quoted by Matt. It is not Matt's own measurement.
Slide 8 was corrected on 2026-08-27 against the official
Claude Code docs. Plan mode blocks edits, not shell commands.
The agent does run commands to explore. The way in is Shift+Tab,
or claude --permission-mode plan at launch. If someone says the
agent does have memory, they are right: Claude Code reads CLAUDE.md
and writes auto memory. The "no memory" frame in this deck is about one thing
only — each session starts with a fresh context window.