Lesson 11 · ChatGPT Mastery Pro+ ~16 min API · Projects · prompting · evals

API, projects, advanced prompting: the power-user end.

The advanced end of ChatGPT, for technical and serious power users. When to graduate from the app to the OpenAI API, how power users live inside Projects, the reasoning patterns that beat default output, and how to build a real eval before you rely on a prompt.

The mental model

Advanced ChatGPT is mostly the API, not the app.

Once you're hitting the app's ceiling — no automation, manual repetition, no version control, no scale — the API is where the serious work happens. The skill stops being prompt-craft and starts being software engineering with AI as a building block. But most people shouldn't leave the app. Find out where you actually stand.

Do it · should you move to the API?

Four signals you've outgrown the app.

Answer honestly. The more yeses, the more the API is worth the engineering overhead.

Answer all four to see where you land.

Prompts that beat the default

Three reasoning patterns.

Default ChatGPT gives default-quality output. These three lift complex tasks 30–50%.

Chain-of-thought

"Think step by step before you answer."

Use for multi-step logic, math, anything it tends to rush.

Self-critique

"Give an answer. Then critique your own answer. Then revise."

Use for consequential work you'd review anyway.

Costs more tokens — skip it for trivial tasks.

Few-shot examples

"Here are 3 examples of good output: […]. Now do the same for this input."

Use for consistent format and tone at volume.

Your call · match the pattern

Which technique fits?

Two more power-user moves

Projects and evals.

ChatGPT Projects

Bundle conversations, knowledge files, and custom instructions per topic or client, so every chat loads the right context.

Keep 5–10 active; archive the rest. Managing 50 defeats the point.

Build an eval

Define what "good" means, write 20–50 test inputs, run your prompt, grade each output, and compare to the previous version.

LLM-as-judge is fine for format/length, shaky on correctness — calibrate with manual review.

Graduate one workflow to the API

Pick one app workflow that involves repetition or batches. Rebuild it via the OpenAI API in 30–60 minutes, add a simple eval, and compare quality and cost to the app version. Set per-call max_tokens, cache repeat calls, and use a cheaper model (like gpt-5-mini) where flagship reasoning isn't needed.

Try a prompt in the playground → Back to ChatGPT Mastery

What you can do now

Recognize the four signals that you've outgrown the app
Make your first OpenAI API call and graduate to real scripts
Use Projects to keep topical work in persistent context
Apply chain-of-thought, self-critique, and few-shot when accuracy matters
Build an eval before relying on a prompt in production

Pro+

Up next in ChatGPT Mastery

Lesson 12 · Custom GPTs — build agents that match your work

Package a system prompt, knowledge files, and tool actions into a reusable agent that knows your context, your tone, and your data. Start lesson 12 →