✈️ The IDE is the fighter jet, the model is the pilot
First, an analogy that brings it all together: the IDE (AntiGravity, Cursor) is the hunt — the machine where you take flight. The model (Claude Code, Gemini) is the pilot — the brain that makes the decisions. And the key point: you switch the pilot based on the mission's goal.
🧭 The core idea
Separate the tool (the hunter) of the brain (the pilot) is liberating: your skill isn’t tied to one product. If one model is better at design and another at reasoning, you choose the right pilot for the flight.
- •O hunt doesn’t decide anything on its own—it gives you the tools and the cockpit.
- •O pilot reasons, writes, and fixes—and can be swapped out at any time.
- •No pilot takes off without a checklist — that’s exactly what the CODA framework is.
the IDE
the model
by objective
= CODA
📝 C — Configure (claude.md)
The first step of CODA is Configure: create a context file — claude.md (or Gemini.md)—in EVERY project. It’s the backbone that the pilot reads before touching anything.
Tech stack & directories
Which technologies the project uses and where the key folders are. This way, the AI won't have to guess the structure.
Commands
How to run, build, and test the project. The pilot needs to know the commands to run and verify it.
Success criteria + how to test
The heart of the file: what “done” means and how to prove it. Without this, everything else loses direction.
💡 Living document
claude.md isn’t written once and forgotten. Whenever the project changes direction, update the file. It’s the shared memory between you and the copilot.
technologies
run and test
what “done” means
always up to date
🗺️ O — Outline (Plan Mode)
The second step is Outline: plan before coding. This is where the professionals’ secret lies — spending ~95% of the time defining the problem and success BEFOREHAND writing a single line of code.
✓ Do it in Plan Mode
- ✓Define the problem in detail
- ✓Go back and forth with the model
- ✓Ask clarifying questions
- ✓Use “ultrathink” for deep reasoning
✗ Avoid
- ✗Say "build X" and start coding
- ✗Skip defining success
- ✗Accepting the AI's first interpretation
- ✗Discover the requirement only at the end
In Plan Mode, the model doesn’t write code: it talks things through, raises questions, and helps draft the specification. When the plan is solid, implementation becomes almost a detail—because the problem has already been solved on paper.
in the plan
round trip
clarifying
deep reasoning
⚡ D — Deploy (parallel agents)
The third step is Deploy: get the agents flying. The productivity boost comes from running multiple agents at the same time (5 to 7) — through the terminal or the IDE itself.
🛫 Agent squad
Instead of waiting for one agent to finish before starting the next, you launch a squad. While one works on the front end, another adjusts the back end, and another writes tests — all in parallel.
You can even use agents on the web connected to a repository, still running even with the laptop closed. The work continues without you.
| Where to run | How | Advantage |
|---|---|---|
| Terminal | multiple sessions | direct control |
| IDE | parallel panels | everything in one place |
| Web + repo | in the cloud | runs with the laptop closed |
agents working together
or IDE
connected to the repo
multiplies
🧩 A — Automation (subagents)
The fourth step is Automation: create specialized subagents, one per subtask. Each subagent has the your own context window, so everything doesn’t spill into a single conversation.
🧠 Why divide
Each model has a limit to how much it can “remember” at once (the context window). By assigning a subtask to a subagent, you keep each conversation concise and focused — and nothing important gets pushed out of memory.
1 subagent
unique to each person
without noise
memory under control
🔎 R — Review & Verify
The last step is Review & verify: the landing. You provide explicit success criteria (written in claude.md) so that "done" stops being a matter of opinion and becomes an objective test.
✅ Success checklist
- ✓Runs no errors
- ✓The tests pass
- ✓Without lint warnings
- ✓Does what was requested
- ✓The changes are explainable
🕵️ Skeptical review
Ask the AI to review its own work as a "highly skeptical senior developer". This framing forces the model to look for gaps instead of praising—and reveals what would break before the user finds out.
runs cleanly
pass
you understand
hunts for gaps
🪶 Less is more
An important counterpoint to CODA as a whole: sometimes a short prompt works better than a super-detailed one. Modern models are good at inferring—and too many instructions can confuse them.
✓ Try it first
- ✓The short, direct prompt
- ✓Let AI infer the details
- ✓Only add details if it gets it wrong
✗ Be careful with
- ✗Filling the prompt “just in case”
- ✗Contradictory instructions
- ✗Micromanage every step
⚖️ The balance
CODA provides structure where it matters (context, plan, verification). “Less is more” reminds you that simplicity is a strength in everyday prompts. Start simple; add detail only when the AI asks or makes a mistake.
first
let the AI
only if you make a mistake
saves
🔬 Practical example
Let’s put together a claude.md for a simple project—a site that shows the weather in 3 cities—with success criteria and tests already defined. This is the “C” in CODA in practice.
# Clima de 3 cidades ## Stack - HTML + CSS + JS puro (sem build) - API: open-meteo (sem chave) ## Diretórios - index.html → página única - /assets → css e js ## Comandos - abrir: é só abrir index.html no navegador - testar: rodar manualmente as 3 cidades ## Critério de sucesso - Mostra temperatura atual de São Paulo, Rio e Salvador - Carrega em < 2s, sem erros no console - Funciona offline depois do 1º carregamento (cache) ## Como testar 1. Abrir index.html 2. Conferir as 3 temperaturas aparecendo 3. Abrir o console: não pode ter erro vermelho
Notice: the pilot (model) now knows what to build, how to run it, and — most importantly — how to know when it’s done.
🏋️ Exercise
Time to take the helm. Complete the two steps below in one of your own projects (it can be small):
- Write a claude.md with success criteria clear and the “how to test” section.
- Run 3 agents in parallel into different subtasks (e.g., one for the interface, one for the logic, one for testing).
At the end, ask one of them to review it as a “skeptical senior developer” and note what they found.
⚡ Ready-made prompts
Copy, paste into your AI assistant, and adapt to your use case.
Atualize o claude.md com o escopo abaixo do projeto e me faça perguntas para preencher os detalhes que faltam (stack, diretórios, comandos, critério de sucesso e como testar). Escopo: [descreva o projeto em 2-3 frases]
(Plan Mode) Quero construir [X]. Me ajude a definir o problema em detalhe e criar as especificações. Faça perguntas clarificadoras antes de escrever qualquer código. Não comece a implementar até o plano estar fechado.
Review this code as a highly skeptical senior developer: - Does it run without errors? - Do the tests pass? Are there uncovered cases? - What would break in production? List the problems by severity and propose a fix for each one.
✅ Module summary
Next module:
5.2 — Skills & Prompts Library