Level 5 — Soft skills for working with AI
This level isn’t about tech. It’s about how to command an AI well. Practice it from day one, in parallel with everything else.
Decompose the task
Section titled “Decompose the task”AI is bad at “build me Uber.” It’s good at “add a login screen with email.” Cutting a big thing into small steps is your number-one skill.
Write a precise brief (prompting)
Section titled “Write a precise brief (prompting)”The more specific the request and context, the better the result. “It doesn’t work” is useless; “clicking Buy turns the page white, and the console shows error X” is gold.
Read errors
Section titled “Read errors”You don’t need to understand an error — you need to copy it in full and hand it to the AI. That’s the basic loop. Red text isn’t a catastrophe; it’s information.
Give the AI context
Section titled “Give the AI context”AI has a limited context window — it can “forget” earlier parts of a long conversation. Help it: paste the relevant file, a screenshot, or the full error text instead of describing things vaguely.
Commit before you prompt (checkpoints)
Section titled “Commit before you prompt (checkpoints)”Get in the habit of committing before a big AI session (Level 2). Then if it goes sideways, you roll back to a clean checkpoint instead of untangling a mess.
Healthy skepticism — verify
Section titled “Healthy skepticism — verify”AI lies confidently (it “hallucinates” — that’s a property, not a bug). Always check that something actually works, not that it looks like it works. Never blindly run commands that delete files or touch production.
Security and money
Section titled “Security and money”Know where a key could leak, where API/server bills can pile up, and what must never be committed. One forgotten key is a real loss.
Know the boundaries
Section titled “Know the boundaries”Know when to call a real developer: payments, personal data, medicine/finance — anything where the cost of a mistake is high.
- Prompt engineering overview (Anthropic) — How to give an AI clear, specific instructions and context.
- Debugging by reading errors (MDN) — Error messages are information, not catastrophe — how to read them.
- 6 steps before taking your vibe-coded app to production (Convex) — The best description of the 'boring 20%': auth, DB, payments, deploy.
A tiny exercise
Section titled “A tiny exercise”Take a feature you want and write it out as 3–5 small steps before touching the AI. Then prompt only the first step. Notice how much better the result is than asking for the whole thing at once.