🎛️ Meta-commands and mastery
The three commands that operate on all the others — /mode, /chain, and /preset —, the document’s five compound recipes, a command’s YAML structure, precedence rules, and what’s coming in v0.2.
🎚️ /mode — global behavior profiles
Activates a persistent behavior profile for the current run. The following commands inherit the mode without you repeating it: one /mode conservador makes /research cite more sources, /decide ask for more evidence, and /estimate widen the range.
| Mode | What changes in the commands | When to use |
|---|---|---|
| conservative | More verification, wider ranges, recommendations with a Plan B | Irreversible decision, money at stake, external audience |
| exploratory | More alternatives, less filtering, tolerates hypotheses | Project kickoff, brainstorming, /discover, and /future |
| executive | Short output, decision first, details on demand | Operating routine, /brief, /next, /review |
The document’s three modes. v0.2 will include modes named by the team (see section 7).
🎯 Copy and run
/mode conservador /research <tema> /truth /gaps /decide entre <A> e <B>
How to verify: /decide returns a plan B and evidence without you asking — a sign that the mode was inherited. Switch to /mode executivo and run it again: the output gets shorter.
New here? Persistent here means "applies until explicitly replaced" — not "applies forever." The mode lasts for the current run; a new /mode replaces the previous one, and the runtime returns to the default when the session ends.
⛓️ /chain — chaining with output handoff
Runs commands in explicit order, passing the relevant output from one to the next. The inline composition from module 1.1 (/truth /gaps /rank) applies them all to the same input; /chain has /gaps work on what /truth produced—and /rank on what /gaps produced.
✓ A real chain
- ✓
/chain(/research -> /truth -> /pushback -> /decide): the pushback attacks only what survived the truth. - ✓Each step's output is visible—you can audit where the conclusion came from.
✗ Decorative chain
- ✗
/chain(/research -> /brief)when /research /brief inline would do the same thing—there is no real dependency between the steps. - ✗A chain of 9 commands where no one can say what step 5 received.
🎯 Copy and run
/chain(/market -> /customer -> /gaps -> /offer) para <ideia>; mostre a saída de cada etapa antes de passar para a próxima e diga o que cada etapa consumiu da anterior
How to verify: the “what it consumed from the previous step” section isn’t empty at any stage — if it is, that stage didn’t need to be in the chain.
💡 Practical tip
Long chains are tiring to read. Above 5 commands, split them into two chains with a /brief in the middle—or promote the chain to a /preset (next topic).
📦 /preset — official command bundles
Expands a short name for a predefined chain of commands with parameters. It’s /chain that got a name: /preset decisao-confiavel instead of typing six commands — and, more importantly, instead of each team member typing six different commands.
🏷️ When a chain deserves to become a preset
- •Repetition: the same sequence ran 3+ times in the last month.
- •Stable parameters: the same criteria, the same level, the same audience — changing only the input.
- •Omission risk: someone already skipped a step and regretted it (the /premortem nobody ran).
🎯 Copy and run
/preset decisao-confiavel sobre <decisão> # equivale a: /chain(/research -> /truth -> /gaps -> /pushback -> /rank criteria=impacto,custo,risco -> /decide)
How to verify: the preset output is identical to the expanded chain’s output — if it differs, the preset isn’t defined the way you thought. Ask for "/preset decisao-confiavel --expandir" to see the chain.
New here? Official preset is a preset defined and versioned by the language (or the team), with a reserved name — unlike a personal shortcut. v0.1 lists five candidate recipes for official presets; v0.2 promises to formalize them (topic 7).
📚 The document’s 5 compound recipes
The document ends with five recommended chains. They aren’t examples — they’re the map of the entire course: each one spans three or four tracks and shows the order in which the families interact.
Recommended compositions — INEMA Command Language v0.1
# Decisão confiável /research /truth /gaps /pushback /rank /decide # Construção de produto /market /customer /gaps /offer /MVP /prototype /eval # Arquitetura de agente /agent /tools /memory /guardrails /eval /integrate # Execução de projeto /blueprint /dependencies /priority /parallel /execute /monitor /review # Tese de negócio /truth /market /pushback /businessmodel /moat /growth /bet
| Recipe | Tracks that cross | The hidden pattern |
|---|---|---|
| Reliable decision | T2 → T1 → T3 | Verify before attacking, attack before deciding |
| Product development | T5.1 → T2 → T5.2 | Market before offer, offer before code, eval at the end |
| Agent architecture | T5.2 | Evaluate before integrating — controlled before connected |
| Project execution | T4 | Plan → order → action → monitor/review loop |
| Business thesis | T1 → T5 → T3 → T5 → T3 | Facts, market, attack, model, bet with an exit |
Notice: /pushback appears twice and /eval appears twice. Attack and evaluation are the steps people skip—the document builds them in so no one skips them.
🎯 Copy and run
/chain(/truth -> /market -> /pushback -> /businessmodel -> /moat -> /growth -> /bet) para <ideia de negócio>; ao final, diga qual etapa mais mudou a conclusão
How to verify: the final response differs from what you thought before running it — if it’s the same, one of the steps (almost always /pushback) was performed too compliantly.
🧬 Internal YAML structure of a command
Under each bar is a definition. The document suggests a nine-field YAML structure for any command — and it’s what makes the language executable by a runtime, not just readable by humans.
Suggested internal structure — example: /truth
name: /truth family: verification version: 0.1 purpose: separar fatos, inferências, hipóteses e desconhecidos inputs: [query, context] tools: [search:auto] rules: - verify unstable claims - never present inference as verified fact output_schema: - verified_facts - inferences - unknowns - confidence acceptance: - critical claims are supported fallback: - state what could not be verified
🔍 The three fields that distinguish a command from a prompt
- •output_schema: the output format is fixed — so it can be tested, chained, and fed into another command.
- •acceptance: declared “done” criterion — the runtime knows when to retry or escalate (Module 1.1, topic 7).
- •fallback: what happens when it can’t — "say what you couldn’t verify" instead of making something up.
🎯 Copy and run
/spec no formato YAML acima para um comando novo: /<nome> da família <família>; preencha os 9 campos, com output_schema de no máximo 5 itens e um fallback honesto
How to verify: the fallback is never "try again"; it’s an observable behavior ("declare X unverified," "escalate to a human").
New here? YAML is a text format for describing structured data, readable by both humans and machines — key: value, lists with hyphens, indentation as hierarchy. It’s the natural format for defining commands because an agent can read the same file a person reviews.
⚖️ Precedence and conflict resolution
What happens when /mode conservador is active, you call one /preset that embeds an exploratory mode, and within it a command has an explicit parameter? The answer in v0.1 is just one rule: the most specific wins.
Illustrative: concentric rings—the inner one (explicit parameter) always overrides the outer one (global mode).
✓ Conflict resolved
- ✓
/mode conservador+/estimate faixa=estreita→ narrow range: an explicit parameter beats the mode. - ✓Preset defines
criteria=impacto,custo; you passcriteria=risco→ risk: the call overrides the preset.
✗ Silent conflict
- ✗Two /mode commands in the same session without knowing which one is active — the last one wins, but no one remembers which it was.
- ✗A preset that changes the global mode without warning: you leave the preset, but /mode remains changed.
🎯 Copy and run
/mode conservador /preset exploracao-rapida sobre <tema> /estimate para <tarefa> faixa=estreita # depois pergunte: qual modo está ativo agora e por quê?
How to verify: the response says which precedence rule decided each case — and states that /mode conservative took effect again after the preset (or says it didn’t, which is a v0.1 bug documented in topic 7).
⚠️ Formal precedence does not yet exist in v0.1
The “specific wins” rule is the recommended convention, not a tested specification. v0.2 promises formal precedence (next topic). Until then, when the result surprises you, ask the runtime which mode and parameters were active—don’t assume.
🛣️ Roadmap v0.2 — what comes next
The document ends honestly: v0.1 is a language operational, not formal. Eight items are listed for v0.2—and each one addresses a weakness you’ve probably already felt in this course.
| v0.2 item | Problem it solves | Where you felt |
|---|---|---|
| Aliases | /SOP vs. /sop vs. /procedimento — the same command with different names | Module 4.1 |
| Parameter types | criteria= accepts any text; nothing is validated | Module 3.2 |
| Output schemas | The /truth output changes shape between runs—breaking the /chain | Module 6.2, topic 2 |
| Formal precedence | "Specific wins" is a convention, not a tested rule | Module 6.2, topic 6 |
| Permission levels | /execute and /automate can act without explicit authorization | Module 5.2 (/guardrails) |
| Tool policies | tools: [search:auto] doesn’t say when NOT to search | Module 5.2 (/tools) |
| Official presets | The 5 recipes are still text, not invokable packages | Module 6.2, topic 3 |
| Conformance tests | There’s no way to prove that a runtime implements /truth correctly | Module 1.1, topic 7 |
Each line in v0.2 names a weakness in v0.1—the document hides none.
Use v0.1 as is
The language works today with any model: commands are prompt conventions. The five recipes and three meta-commands already improve the quality of what you get.
Document where it failed you
Whenever a command returns something unexpected, note the command, parameters, active mode, and output. That’s the material for v0.2.
Propose in YAML
A new command or a correction is more useful as 9-field YAML (topic 5) than as a suggestion in prose.
🎯 Copy and run
/review do meu uso da INEMA Command Language neste curso: quais comandos usei mais, quais nunca usei, onde a saída surpreendeu e qual item da v0.2 resolveria cada surpresa
How to verify: the output lists at least one command you’ve never run — and it’s worth trying before you close the tab.
💡 Where to go now
You finished the course. If you revisit just one page, return to module 1.1 (the syntax and the 10 rules) and this 6.2 — the beginning and the end are the two places where the entire language fits on one screen.
Quick self-check (optional): what's the difference between inline composition (/truth /gaps /rank) and /chain(/truth -> /gaps -> /rank)?
🎓 Module Summary
You completed the course:
Track 6 complete—and with it, all 6 tracks and the 100 commands of INEMA Command Language. Return to the beginning to review the complete map or open "My Journey" to see what you marked as a question.