Guide

Intent Engineering

Intent Engineering is the first lesson because it names the failure mode.

Source: docs/intent-engineering.md

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.

VersionIntent
WeakUse an agent to clean up notifications.
BetterMake 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