What I Actually Want You to Leave With

A cheat sheet for the agentic track. The short list.

How to Use This Sheet

One page. If I only had five minutes with you in office hours at the end of the semester, these are the things I would say. Not a recap of the readings. Not a checklist. The few dozen heuristics that you will use for the rest of your career, because the specific tools will keep changing but these will not.

How to Think With AI

An LLM is the world’s best autocomplete, not a thinker. It predicts the next token given everything that came before. Plausible is not the same as true. Every time you feel tempted to trust the output without checking, remember this sentence.

Tokens are not words. Roughly four characters each. You pay per token, your context window is measured in tokens, and code eats more tokens than prose. When your bill or your context is under pressure, the fix is almost always to send less, not to send better.

An agent is an LLM plus tools plus a loop. Think, Act, Observe, repeat. Every framework, every “agentic” feature, every bells-and-whistles platform sits on top of that. Learn the loop once and the new framework you meet next year is easy.

Hallucinations are not a bug. They are the default behavior of prediction under uncertainty. Give the model more context, ask it to flag uncertainty, ask it to show its reasoning. Those three moves are the bulk of what “bias toward truth” means in practice.

Record the meeting. Feed the transcript to AI. Perfect recall plus synthesis across meetings is one of the highest-ROI skills you will develop this semester. Five minutes of setup, lifetime of payoff. Announce recording and get consent.

How to Plan With AI

PRD is immutable. Plan and roadmap are living documents. The PRD is the source of truth. It gets “DO NOT MODIFY” at the top once it is locked. The plan tells a human how to think about the work. The roadmap tells the AI what to do next.

Pipeline: PRD, then research, then plan, then roadmap. Each one is a different document because it answers a different question. Collapsing them into one is how you end up with a plan built on assumptions instead of facts.

Two folders, not one. aiDocs/ is tracked. ai/ is gitignored. Permanent project knowledge (context.md, PRD, architecture, coding style) goes in aiDocs/. Working process artifacts (roadmaps, research, notes) go in ai/. claude.md and .cursorrules are personal config; gitignore them too.

context.md is the most important file in your repo. Bookshelf, not textbook. One or two sentences per referenced doc. AI reads on demand; you do not need to duplicate content.

Be ruthless about MVP scope. You can always add later. Pressure-test against time, people, and the original problem. If the new feature does not trace back to the PRD, it does not belong in this sprint.

Run the frenemy prompt before you commit to a plan. “Assume I’m wrong and make me prove otherwise.” Then take the output to a fresh session and decide what to act on. Build up, stress-test, refine.

How to Build With AI

Implement one phase at a time. Do not hand over the whole roadmap. Phase by phase keeps errors small and reversible.

If AI can run a command, AI can test your app. No CLI surface means no autonomous loop. JSON on stdout, plain text on stderr, proper exit codes, --help. Four rules. Break one and the loop stalls.

If AI can see what happened, AI can fix it. Structured JSON logs, not prose. Log function entry, function exit, errors with context, external calls, DB queries. Pino or structlog, not console.log.

Test credentials go in .testEnvVars, not .env. Explicit separation between the app and testing. AI sources the test file on its own.

Never paste a secret into a prompt. Reference the env var by name. Prompts may be logged. Verify .gitignore before the first commit, not after.

TDD for units. Explore-then-codify for integration. Unit tests catch logic bugs; integration tests catch wiring bugs. Verify Red is really red before you write the implementation.

Trust but verify, and do the verification with a sub-agent. The agent that wrote the code has motivated reasoning. Deploy a fresh one to cross-check the work against the roadmap. No code changes; just report.

When tests fail, let the loop run two or three cycles before you step in. Intervene only on three triggers: same error three or more times, fixes getting more complex instead of simpler, or more tests failing than before.

The confidence trap is real and it is documented. AI makes you faster and more confident. The Stanford finding is that AI-assisted developers ship more vulnerabilities and report higher confidence in their security. The fix is process: sub-agent review, fresh-session PR review, catching the three patterns (SQL injection, hardcoded keys, prompt injection) before they ship.

How to Keep AI Honest

Instruction files are the highest-leverage investment. Write a CLAUDE.md this week. Identity and invariants at the root. Repeatable procedures in ai/instructions/. One-step shortcuts as slash commands. Dynamic, then codified, then automated. Do not automate before you have codified.

Plan in one session, build in a fresh session, review in a third. Multi-session discipline is the defense against context pollution, which is the number one project killer. When a long session starts getting worse, the fix is almost never a better prompt.

The Law of Witnesses: match agent count to risk. One agent for routine work. Two for high-stakes. Three or more for anything truly consequential. Independent agents converging is real signal; one agent’s confidence is noise.

Sub-agents cost three to five times the tokens when controlled, five to ten times when unchecked. Budget for it. Fewer specialized agents beat many generic ones by 40 to 60 percent, published number. If your agent’s prompt does not fit on one screen, it is too big.

The escalation ladder: re-prompt, fresh session, sub-agent review, manual. Four levels. If rephrasing once did not work, rephrasing twice will not work. Golden rule: when in doubt, fresh session.

Put the AI down for five things: deep learning, novel problems, security-critical code, stuck loops, and team alignment moments. Not everything should be done with an AI in the loop. Knowing when to close the laptop is part of the skill.

The Shortest Possible Summary

You are the forklift driver. The AI is the forklift. The warehouse is the system you build around both: instruction files, two-folder structure, CLI surface, structured logs, multi-session discipline, and a verification habit. Every principle on this page is a piece of that warehouse. You do not need all of them on day one, but the semester ends with all of them in place, because the project you are building only works if the warehouse does.

Now go build something.