Intent inference and alignment for persistent AI agents.
Reduce your human's cognitive load. They throw raw thoughts, the agent infers the task — and verifies it's the right task before executing.
When humans direct AI agents, there's a constant gap between what they say and what they mean. The task is "do A." The intention is why — what outcome they're actually driving toward. A is just one of many possible paths.
Most agents execute the literal task. Good agents understand the intention and execute toward it.
The Intention Engine gives your agent a protocol for:
- Gap classification — distinguish spec gaps (unclear how) from intention gaps (unclear why). Different gaps need different fixes.
- Context-layered inference — stack user goals, topic context, recent memory, project state, and conversational momentum to infer intent.
- Premortem checks — before executing anything expensive or irreversible, ask "what's the most likely way this fails?"
- Quality bar assessment — distinguish "done adequately" from "done well" and match the right bar to the task.
- Negative intent checks — identify what NOT to optimize for, preventing the Klarna trap.
- Wasted work detection — verify the task serves the intention before executing.
- Calibrated push-back — challenge tasks that conflict with stated goals or when better alternatives exist.
This skill focuses on understanding intent and aligning execution. It does not cover:
- How to think about problems — see Activated Thinker for anti-binary thinking, gardener mindset, friction protocol, and capability building
- Behavioral mode detection — see Activated Thinker for crunch vs exploratory mode
These skills complement each other: Intention Engine tells you what to do, Activated Thinker tells you how to approach doing it.
clawhub install mouserider/intention-engineOr copy the skill folder into your OpenClaw workspace's skills/ directory.
This skill is directly inspired by and built upon Nate Skelton's Intent Engineering framework:
- Intent engineering — the distinction between task execution and intention alignment
- Premortem prompting — forcing failure imagination before committing to a plan
- Quality bar distinction — "done adequately" vs "done well"
- Context layering — structured stacking of context for richer inference
- Spec clarity ≠ intention clarity — they fail differently and need different fixes
- The Klarna/$60M case study — the danger of optimizing for stated metrics while destroying unstated constraints
MIT