⌨️ The /inema language
From the first slash to the full runtime. In seven topics, you’ll learn the grammar that makes the same 100 commands work in any agent.
🧠 What is INEMA Command Language
The INEMA Command Language is a concise operational language: a vocabulary of 100 commands that controls how AI models and agents research, verify, analyze, critique, decide, plan, execute, and communicate. You don’t program—you directs.
New here? One DSL (Domain-Specific Language, domain-specific language) is a small language built for a single purpose— like SQL is for databases. Here the purpose is direct AI with precision. One declarative command describes what you want (e.g., “separate fact from inference”) without specifying the internal steps.
💡 The central principle
A command isn’t a magic word: it’s the visible tip of a protocol.
- •
/factcheckdoesn’t "ask for a check" — it breaks the text into claims, classifies each one (confirmed, inaccurate, false, inconclusive, opinionated), and delivers the verdict. - •The same command in different agents produces comparable output format — because the protocol, not the model, defines the form.
⌨️ The base syntax
All the syntax fits on one line: /comando [parâmetros opcionais].
Parameters use the form key=value. No mandatory quotation marks, no semicolons.
Real anatomy of /rank criteria=impact,cost,risk: command → protocol → output with confidence and log.
🎯 Try it now (copy and run)
📜 The 10 language rules
The rules are short, but they solve the problems that break improvised prompts: ambiguity, execution order, and conflicting instructions.
⚖️ Highlight three rules
- 5.Conflict: the more specific rule wins; if tied, the more recent one wins. This is the language’s official precedence.
- 6.Division of labor:
/presetexpands macros,/chaincontrols sequence,/modedefines global behavior. - 9.Generation ≠ verification: when the risk is significant, the person who generates cannot be the person who verifies—separate the roles.
✓ Rules used well
- ✓Treat commands as declarative instructions, not as magic spells.
- ✓Respect the explicit chain order (
/chain(/research -> /truth -> /decide)). - ✓Require acceptance criteria for critical commands.
✗ Errors the rules prevent
- ✗Stacking contradictory commands and hoping that “the model understands.”
- ✗Assuming typing order = execution order outside an explicit chain.
- ✗Accept output without confidence or a record in risk decisions.
💡 Practical tip
Memorize only rule 5 (specific > recent). Look up the other nine — they exist precisely so you don’t have to memorize them.
🔗 Inline composition
Several commands on the same line form a instruction package: /truth /gaps /pushback /rank.
It's the fastest way to build a robust response without ceremony.
🎯 Copy and run
/truth /gaps /pushback /rank criteria=impacto,custo,risco sobre
How to verify: the response must include all four hallmarks — separating facts from inferences, identifying gaps, challenging the idea, and ranking options.
💡 Practical tip
Start with 2–3 commands, not ten. A giant package dilutes the model’s attention; a good composition combines complementary families (truth + critique + decision).
⛓️ Chain with /chain
When the order matters and the output of one step feeds the next, use an explicit chain:
/chain(/research -> /truth -> /rank -> /decide).
New here? Think of a pipeline: a pipeline where each station receives the work from the previous one.
🎯 Copy and run
/chain(/research -> /truth -> /rank -> /decide) sobre
How to verify: each step must cite where its input came from (the previous step) — without that, no output was handed off.
⚠️ Attention
An overly long chain hides errors for too long. If step 1 is wrong, every step after it inherits the error. For long workflows, alternate /verify or /review like control valves.
🎛️ /mode and /preset
They’re the two highest-level controls. /mode defines persistent global behavior
(conservative, exploratory, executive). /preset expands a short name
into a predefined chain of commands with parameters — a macro (a shortcut that expands into multiple instructions).
A preset is a box within a box: when expanded, it reveals a complete chain of commands.
💡 Practical tip
Use /mode conservador on regulatory or medical topics; /mode exploratorio in brainstorming. The mode changes the risk appetite of all subsequent commands.
🧾 Runtime: registration, acceptance, fallback
The execution contract has three parts: the the runtime logs everything (inputs, outputs, tools, confidence, failures), every command can have acceptance criteria (how to know it worked) and a fallback (what to do when it doesn’t work).
Operational record
Language Rule 7: inputs, outputs, tools used, confidence, and failures are recorded. Without a log, there’s no audit.
Acceptance criteria
Rule 10: define in advance what counts as success (“every critical claim is supported”). Prevents accepting a polished but empty answer.
Safe fallback
When it can’t verify, the command states what it couldn’t verify—it never makes anything up. Failing honestly is the correct behavior.
🧪 How a command describes its own contract (internal YAML)
name: /truth family: verification purpose: separar fatos, inferências, hipóteses e desconhecidos acceptance: - critical claims are supported fallback: - state what could not be verified
How to read: acceptance is the success test; fallback is honest failure behavior. Module 6.2 lays out this entire structure.
Quick self-check (optional): two commands conflict in one line. Which rule resolves the conflict?
🎓 Module Summary
Next Module:
1.2 - Truth and verification — the 10 commands that keep AI from presenting guesses as facts.