Intent Engineering
Intent Engineering is the first lesson because it names the failure mode.
Fast AI iteration feels productive. Without checkpoints, it becomes motion without steering: unvetted advice, local hacks, hidden assumptions, and fixes that make the next fix harder.
The rhythm is: clarify intent, burst for breadth, pause and reflect, structure the work, then iterate.
Clarify intent
Write one sentence before touching code.
| Version | Intent |
|---|---|
| Weak | Use an agent to clean up notifications. |
| Better | Make future notification changes safer by moving duplicate prevention to the boundary where sends happen. |
The sentence should name the outcome, not the activity.
Burst
Use the model for cheap breadth:
- likely causes;
- plausible solution levels;
- risks;
- files to inspect;
- checks that would fail the wrong patch.
Do not commit during the burst. Generate, then pause.
Pause and reflect
Ask:
- Is this necessary for the intent?
- Is it sufficient?
- What assumption does it depend on?
- What would prove it wrong?
- What work is tempting because it is nearby, not because it matters?
This is where /problem-space, /problem-statement, and /solution-space enter.
Structured pass
Convert the best current understanding into artifacts:
- aim;
- problem-space map;
- problem statement;
- solution-space comparison;
- evidence checklist;
- agent brief.
The artifact matters because another skill, subagent, or future session can inherit intent without re-inferring it from chat history.
Iterate
After execution, do not ask only “did it work?” Ask:
- What evidence says it worked?
- What did review find?
- What did dissent challenge?
- What should become a skill, subagent, guardrail, metis, signal, or salvage note?
Intent Engineering keeps speed useful by making the control system explicit.
Exercise
Before opening the codebase, write down: intent, burst topics, pause questions, structured artifacts to produce, and what would prove this was the wrong task.
Go deeper
- Intent Engineering — the source essay for clarify, burst, pause, structure, iterate.
- Alignment Is the Constraint — why intent must carry mechanism, feedback, and guardrails.
docs/model-fit.md— the next skill: shape the ask around model strengths before constructing context.docs/context-construction.md— turning intent and model-fit into usable context.
Navigation
- Previous: Tutorial
- Up: Docs Home / Curriculum
- Next: Model-Fit Framing