Climb the ladder’s seven steps
One via it’s the path the agent uses to reach a system: request data, read a file, click a screen. The runtime/LEIA-ME.md organize the paths into a seven-step ladder levels.
You start at level 1 and work your way down until you find the first method that exists and works. The kit's wording: use the highest available.
🆕 New here? The ladder's words
- API — the official door a system opens so other programs can request data. E.g., "give me October's sales."
- MCP — a standard for providing tools to agents. An MCP server says “I have tool X,” and the agent can then call it by name.
- CLI — a program used by terminal commands, such as
codex execorgit. - SDK — a ready-made library (npm or pip package) that a programmer uses to interact with the system.
- Local bridge — use what the system leaves on your computer: a CSV export, a folder, a SQLite database, a local port.
How to read the diagram: the number is the order in which you test. The color shows stability: green is high, blue is medium, red is low. Notice that 6 is blue and 5 is red: being lower down doesn't mean it's worse. Topic 2 explains.
| Level | Via | Example | Stability |
|---|---|---|---|
| 1 | Official API | ERP API | high |
| 2 | MCP | claude mcp list | high |
| 3 | CLI | codex exec, gh, git | high |
| 4 | SDK / library | npm/pip package | average |
| 5 | Computer use | automated browser, clicks on the screen | downloads |
| 6 | Local bridge | CSV export, folder, SQLite database, local port | average |
| 7 | Reverse engineering | observe the app running to find the path | lab only: breaks in the next update |
What to look for in the table: it’s the table from the runtime/LEIA-ME.md, without changing a single word. The “Example” column shows the kind of thing found at each level; you don’t need to run these examples now.
path to the system
test order
the three solid ones
exported file
Understand why stability comes first
Stability is how well the route keeps working when the other side changes. An API has a contract: the provider gives notice before making changes. On a website, a button moves, and the robot that clicked there gets lost.
The kit's README sums up the idea: the agent uses always the most stable one available. That's why the local bridge (6, medium) beats computer use (5, low) when both are available. Recipe R5 says the same thing: use the browser only when there is no API, MCP, CLI, or export.
✓ Stable method
- ✓ It comes with a contract: the manufacturer maintains it
- ✓ Changes come with a notice and version
- ✓ You configure it once and forget about it
- ✓ When an error occurs, it’s clear
✗ Fragile approach
- ✗ Depends on where the button is on the screen
- ✗ Breaks without warning in an update
- ✗ Needs fixing every week
- ✗ Sometimes fails silently
💡 The real cost is maintenance
The stable route can sometimes take more work on the first day: finding the export, configuring the bridge. But what matters is the following month. The route that breaks less is the one that costs less to maintain.
can it handle change?
the manufacturer maintains
the cost of the following month
breaks without warning
Require a test before saying "it can’t be done"
The agent says "it can't be done" too soon. It couldn't find documentation and concludes that it doesn't exist. The kit has a rule against this, written in the LEIA-ME.md and repeated in the AGENTS.md.
The rule: before concluding that something is impossible, the agent test each level with a command. “I couldn’t find it” doesn’t count. “I ran this and got this” does.
Rule: if the agent says "can't be done," ask it to **test** each level with a command before concluding. Often the route exists and simply wasn't documented.
runtime/LEIA-ME.md. O AGENTS.md asks the agent to do it on its own.How to read the diagram: The red loop returns to the left with the next step. Each pass leaves a recorded piece of evidence. An honest "can't be done" comes with this list; a "can't be done" without a list is a guess.
⚠️ Prematurely calling something "impossible" is costly
When the agent can't find documentation, it tends to say that the program can't be read externally. Often, testing level by level reveals an export option hidden in a menu. If you accept the first "can't be done," you resort to a workaround unnecessarily.
"I didn’t find it" ≠ "it doesn’t exist"
one command per level
the output on screen
the rule is already there
Climb the ladder with Sônia’s ERP
Sônia wants total sales by customer from the distributor. The ERP is old: it has no API. What it can do is export a sales CSV, o runtime/exemplos/erp-vendas.csv from the kit.
Look at the ladder one step at a time. This is the most common case in a small office: an old system that only exports a file.
| Step | Does it exist for the ERP? | Decision |
|---|---|---|
| 1 · API | no: the manufacturer doesn’t offer it | goes down |
| 2 · MCP | there’s no ERP MCP server | goes down |
| 3 · CLI | has no terminal command | goes down |
| 4 · SDK | has no library | goes down |
| 5 · computer | you could click through the screens… | it exists, but confidence is low: save it and check the next one |
| 6 · local bridge | yes: export a sales CSV | stop here: average beats low |
What to look for in the table: at step 5, the answer is "exists," and the ladder keeps going anyway. That’s the rule in topic 2: R5 only accepts the browser when there’s no export, and there is one here.
How to read the diagram: the top line is the chosen path: the ERP outputs the file, and the bridge provides a ready-to-use tool. The crossed-out red box below is the discarded path. In track 2 (recipe R3), that bridge becomes a real MCP tool.
what the ERP already does
erp-vendas.csv
level 6, average
N4: reads without asking
Climb the ladder with Clara’s calendar
Clara’s clinic schedule is a spreadsheet. A spreadsheet doesn’t have an API, MCP, CLI, or SDK: it’s a file. So the ladder goes down to the local bridge, the same as Sônia’s. In the kit, it’s the runtime/exemplos/agenda.csv, with Dr. Ana and Dr. Bruno’s availability.
The difference shows up later. Sônia only wants to read. Clara, one day, will want the agent to mark a time. And the same route can have different policies for reading and changing.
🆕 New here? N2 and N4
These are autonomy levels for runtime/POLITICA.md. N4: the agent acts on its own, without notice (suitable for reading). N2: the agent acts only after asking each time. The full scale, from N0 to N4, is in module 4.1.
✓ Read the calendar (N4)
- ✓ "What times are available on the 7th?"
- ✓ The agent reads the file and responds
- ✓ Nothing changes in the spreadsheet
- ✓ Can run without asking
✗ Change the schedule without asking
- ✗ Schedule a patient on its own
- ✗ Switch from “free” to “busy” without notice
- ✗ An error becomes a patient with no appointment
- ✗ That’s why changes are N2: ask first
| Agenda da clínica (planilha) | Ponte local (arquivo) | 6 | ler/escrever agenda.csv | alterar (N2) | pendente |
💡 A tip for Clara
Start with read-only access. Once you trust its answers about available times, add write access, always asking first. It’s the same file; the policy is what changes.
Dr. Ana and Dr. Bruno
N4
N2, asks first
different policies
Climb the ladder with your system
Now it's your turn. Choose one work system: the store’s software, the supplier portal, the inventory spreadsheet. Ask the agent to climb the ladder with you, testing each level.
The prompt below uses only files from the kit. It asks for the test at each level and forbids changes to the system: at this stage, the agent only investigates.
Open claude in the kit folder and paste (replace what’s between < >):
Leia runtime/LEIA-ME.md. Quero conectar o <nome do sistema>, que eu uso para <o que você faz nele>. Suba a escada das vias comigo, um nível por vez. Em cada nível, me diga como testar com um comando, mostre o resultado e só então passe ao próximo. Não conclua "não dá" sem testar todos. Não altere nada no sistema. No fim, me diga qual é a via mais estável que existe e por quê.
One system at a time
Mixing three systems in one request muddles the evidence. Do one, take notes, then move to the next.
Answer what only you know
The agent can ask whether there's an "export" menu or an integrations site. Check the program and report back.
Save the choice
The route you choose becomes a row in the CAPACIDADES.md in the next module, and the foundation of your final project.
Quick test (optional): Sônia's ERP can be operated through the screen (level 5) and also exports CSV (level 6). Which route should she use?
by order
level by level
at each step
the most stable
🎓 Module summary
Next module:
1.4 — The capabilities map