AI Training
Write instructions AI can actually follow.
prompt engineering at MMP means writing instructions as verifiable contracts — objective, protected values, constraints, acceptance — not magic phrases.
What it is and isn’t.
What a production prompt IS
- A testable contract
- Objective
- Protected values
- Constraints
- Verify step
- Acceptance
What a production prompt ISN’T
- Role-play fluff
- Magic phrases like “think step by step”
- Vibe-based instructions with no acceptance test
Anatomy of a production prompt.
Prompt skeleton
# objective 1objective: rewrite 40 product blurbs in brand voice 2protected: [prices, SKUs, model names, claims] 3constraints: no new claims · ≤ 40 words · US spelling 4verify: diff vs source · claim-check · read-grade 5accept_when: all checks pass and owner signs
Each line maps to a check that can pass or fail.
Patterns that hold up.
Single objective
One prompt, one job.
Protected values
List what must never change.
Hard constraints
Format, length, spelling, scope.
Few-shot examples
Two good, one bad.
Output format
Schema, not prose.
Refusal rules
Say what to do when input is unclear.
Failure modes we train out.
- Vague verbs (“make it better”)
- Conflicting constraints
- No acceptance test
- Prompt sprawl past one screen
- Unguarded edits to protected values
- Instructions the model can’t verify
Our loop.
Draft
Contract-style prompt from your brief.
Golden set
20–50 known-good cases.
Evaluate
Automated checks, not vibes.
Tune
Fix prompt or context, not output.
Sign-off
Owner accepts before anything ships.
Ready to write prompts that hold?
Start with your situation. We’ll show you the contract style on your own work.