Understand why quota isn’t dollars
The kit runs through subscription for Claude Code and Codex. You pay for the monthly plan, period. You don’t get billed per call.
But the subscription has a usage limit per period: the quota. Each request to the agent uses up a bit of it. When it runs out, you wait for the period to roll over.
The first line of the runtime/ROTEAMENTO.md summarizes the module: "Don't always use the most powerful model. Each call uses up some of your subscription quota."
🆕 New here? Three words from this module
- Quota — the usage limit on your subscription for a given period. It’s not money: it’s how much more you can ask for before you have to wait.
- Model — the AI's "head" that responds. Claude Code has several (
opus,sonnet,haiku…), Codex too. The more powerful ones think better and use more of your quota. - API equivalent — how much the same task would cost if billed by API usage. It’s a benchmark for comparison, even if you don’t pay that amount.
How to read the diagram: the purple bar is the quota. The short blue blocks are requests to the smaller model; the long red ones are the stronger model used unnecessarily. Two unnecessary calls to the top tier use more than several correctly routed tasks. The sizes are illustrative, not measurements.
✓ Quota (what you have)
- ✓ Comes with the subscription you already pay for
- ✓ It ends, but comes back in the next period
- ✓ Gets better results when each task uses the right model
- ✓ It’s what the entire kit uses
✗ API dollars (what you don’t need)
- ✗ Separate account, billed per use
- ✗ No refunds: what you spent, you paid
- ✗ No kit recipe depends on it
- ✗ In the course, it appears only as a benchmark
💡 A real number to give you an idea
The kit's CHANGELOG records that the three-role team (module 3.2), creating a one-sentence file, spent the equivalent of ~US$ 0,73 of API costs. With a subscription, this is quota, not a charge. But it shows that even a tiny task has a cost when it calls large models.
monthly plan
limit per period
benchmark, not a bill
model by task
Read the four model levels
The kit doesn’t use standalone model names. It uses model level: super, top, executor, and smaller. Each level says when to use and point to the model in Claude Code, Codex, and on your computer.
This is the table from runtime/ROTEAMENTO.md, copied from the kit:
| Level | When | Claude Code | Codex | Local (free, Ollama) |
|---|---|---|---|---|
| super | rare: difficult problem, major decision | fable · high effort | gpt-6-astra · high | — |
| top | plan, decide, review the hard parts | opus · medium | gpt-6-astra · medium | qwen3.8:27b |
| executor | do everyday work | opus · low / sonnet | gpt-6-sol · low | qwen3.8:27b |
| smaller | triage, format, check | haiku / sonnet | gpt-6-luna | llama3.2 |
What to look for in the table: the "When" column is the one you use day to day. The three on the right only tell you which name to type in each tool. Note that gpt-6-luna, the bridge’s default in module 2.1, is the lower level.
🆕 New here? "Effort" and "Local"
Effort is how much the model thinks before responding (low, medium, high). The same model with high effort thinks more and uses more quota. Local is a model that runs on your own computer through Ollama: it uses no quota, but depends on your machine being powerful enough. The doctor shows whether you have Ollama (line ollama, marked as optional).
How to read the diagram: the width of a step represents how often it should appear in your day. The smallest one (purple, glowing) is the broad foundation. The super model, red and narrow, is an exception: if it appears all the time, something is wrong with the routing.
big decision
plan and review
do
triage and check
Start with the smallest thing that solves it
Routing rule 1: "Start with the smallest option that works; move up only if it fails." O AGENTS.md from the kit repeats this for both agents: "use the smallest model that gets the job done".
In practice, you move up one level at a time, and only for a reason: the lower level got something wrong, let something through, or couldn’t handle the size.
Smallest fix: format and check
Clara wants the list of available times formatted as a message. That’s formatting: a lower level.
Did it fail? Executor
Sônia asks for the total per client from the ERP CSV. If the smaller model gets the sum wrong, move up to the executor. She checks the correct total by hand (module 2.3).
Decision? Top
"How do I connect the ERP without an API to the agent?" involves planning and decision-making. That’s where the top tier pays off.
Very: only if the top gets stuck
A truly difficult problem, a big decision. Rare, as the table says.
The Codex bridge is set up this way from the start: without you saying anything, it uses gpt-6-luna, the lower level. To move up, you change the variable CODEX_MODELO only for that call.
In the terminal, from inside the kit folder (module 2.1 bridge, already with chmod +x):
runtime/pontes/codex-exec.sh "Responda apenas PONG"
Result confirmed in CHANGELOG 0.1.0:
PONG
To raise Codex to executor level for this call only, put the variable in front:
CODEX_MODELO=gpt-6-sol runtime/pontes/codex-exec.sh "Responda apenas PONG"
PONG using the standard gpt-6-luna. The variable applies only to the line where it appears; the next call goes back to the lowest level.💡 Move up with a written reason
When you move up a level, write down why in one line ("the kid got the total for Padaria Lua wrong"). After a few weeks, these lines show which tasks should start at a higher level.
rule 1
one step at a time
bridge pattern
switches to a call
Ask another model to review it
Rule 3: "Review by another model (e.g., Claude does it, Codex reviews — recipe R1) catches errors the same model doesn’t see."
It’s the same idea as asking a colleague to reread your writing. The person who wrote it tends to read what they meant to write, not what’s actually there. A model from another company has different blind spots.
How to read the diagram: the blue arrow is the bridge from module 2.1 carrying the request to Codex. The dashed purple curve is the review coming back. Codex only reads: Claude decides what stays, and you have the final say.
Open claude in the kit folder and paste (R1 recipe prompt):
Use runtime/pontes/codex-exec.sh to ask Codex to review the README.md file: what might confuse a beginner? Then compare its feedback with your opinion.
✓ Review by another model
- ✓ Different blind spots cancel each other out
- ✓ The reviewer can be a lower-tier model
- ✓ Disagreement becomes a question for you
- ✓ Already included in the kit’s bridge
✗ The same model reviewing itself
- ✗ Tends to agree with what it just did
- ✗ Repeats the same reading error
- ✗ Gives a false sense of being “verified”
- ✗ Uses up quota without adding a fresh perspective
rule 3
Claude does it, Codex reviews
bridge pattern
in disagreements
Build a team only when it’s worth it
Rule 2: "Use a team of agents only when the work is worth more than the cost of assembling the team." A team (module 3.2) calls several models, and each one reads the task and the project from scratch.
The kit's own example shows the cost: the team created a one-sentence file and spent the equivalent of ~US$ 0,73 in API costs. It was great as a test. For writing one sentence in day-to-day work, a single agent would be enough.
| Task | Is one agent enough? | Is a team worth it? |
|---|---|---|
| Schedule Clara’s free time | Yes, lower level | No |
| Answer a question about the POLICY | Yes | No |
| Sônia’s monthly ERP summary, with verified totals | Risky: no one checks | Yes: plan, do, and check |
| Build a client system bridge | Risky | Yes |
What to look for in the table: A team makes sense when mistakes are costly and someone needs to review the work. If the task fits in one sentence and mistakes are immediately visible, one agent is enough.
⚠️ An agent inside an agent costs more than it seems
Rule 4, copied from the kit: "An agent call from within another agent carries the project's instructions and tools. For simple bridges, use a lightweight folder or --bare from Claude." In other words, each internal call rereads the folder’s rules. In a crowded folder, that multiplies the cost.
💡 Quick question
Before assembling a team, ask: "if the agent gets this wrong, will I notice right away?" If you will, use a single agent. If you won’t, the team’s reviewer is worth the quota.
short task
work greater than the cost
less to reread
R2 equivalent
Keep the table up to date
The last line of the ROTEAMENTO.md warns: "Model names change. Check with claude --help e codex --help and update this table."
The levels stay. What changes is the name inside each cell. That's why the recipes say "lower level" rather than a fixed name.
In the terminal, one command at a time:
claude --help codex --help
Or let the agent compare. Open claude in the kit folder and paste:
Rode claude --help e codex --help, compare com a tabela de runtime/ROTEAMENTO.md e me diga quais nomes de modelo mudaram. Não edite o arquivo: proponha a mudança numa linha da tabela Aprendizado da runtime/POLITICA.md.
ROTEAMENTO.md stays the same until you approve.How to read the diagram: the glowing purple box is you. The agent checks and proposes, but doesn't change the rules file on its own. It's the "propose → approve → incorporate" of the POLITICA.md, which you’ll examine in depth in module 4.2.
Quick test (optional): Sônia only wants to format the total by customer that she has already checked as a table. Which level should she use?
source of the current name
name changes
agent proposes
and only then changes
🎓 Module summary
CODEX_MODELO only if it makes a mistake.--help.Next module:
3.2 — Three-role team