Constraint Stack for AI Work: A Practical Template

If constraints are the moat, we need a repeatable way to define them.

Here is the template I now use before handing work to an AI system.

The Constraint Stack

1. Goal
   What exactly should be true when this is done?
2. Acceptance criteria
   Measurable pass/fail rules.
3. Task verifier
   How we independently check the output is actually correct.
4. Failure boundary
   What the agent is not allowed to touch.
5. Rollback and audit
   How we recover and what evidence we keep.

Worked Example: SQLite Rewrite Prompt

A vague prompt: "rewrite this SQLite layer in Rust."

A constrained prompt:

• Goal: Rust implementation of the same query behavior.
• Acceptance criteria:
  • p99 read latency <= 50ms on benchmark workload
  • handles 10k queries/second on local benchmark harness
  • uses indexed lookup path for primary-key reads
  • zero data-loss regression tests pass
• Task verifier:
  • run benchmark harness and compare against baseline
  • run query planner inspection for PK lookups
  • run integration and regression test suite
• Failure boundary:
  • do not modify schema migration files
  • do not change wire protocol
• Rollback and audit:
  • feature flag for old path
  • benchmark report attached to PR
  • reproducible test command in CI logs

Same model. Different result quality.

Why This Helps

Most AI failures are not model-too-dumb failures. They are objective-too-vague failures.

The constraint stack does two things:

• gives the model a target it can optimize for
• gives humans a way to reject plausible-but-wrong output quickly

That is the difference between moving fast and moving randomly.