Skip to content
>_MyGearHut

12 Rules for AI — Operator Cheatsheet

Twelve prompt-discipline rules for operators who need reliable AI output — context, constraints, verification, and when to stop trusting the model.

Why prompt discipline beats prompt magic

Most AI failures are operator errors: vague goals, missing context, no success criteria, and no verification step. These twelve rules are the minimum bar MyGearHut uses before any prompt goes into production, a client deliverable, or a recurring automation.

Print this. Pin it next to your chat window. Run every serious task through it once.

The 12 rules

1. Name the job, not the vibe

Bad: "Make this better."
Good: "Rewrite this product page for HVAC owners — 120 words, CTA to book a call, no jargon."

Action: Start every prompt with role + task + output format.

2. Give constraints before creativity

List what must stay true: brand voice, legal claims, word count, banned phrases, target reader, deadline.

Copy-paste opener:

Constraints (do not violate):
- Audience: [who]
- Tone: [plain / technical / sales]
- Must include: [facts, links, numbers]
- Must NOT: [invent stats, mention competitors, etc.]

3. Attach source material

If the answer must reflect your data, paste it. Summaries beat links. For code, paste the file or diff — do not describe it from memory.

4. Ask for structure, not prose

Tables, numbered steps, JSON, checklists, and bullet tiers reduce hallucination and make output usable without editing.

5. One task per thread

Multi-task prompts produce multi-task garbage. Split research, drafting, and formatting across turns or separate chats.

6. Plan before execute (for anything that ships)

For code, campaigns, or client work: request a plan first. Approve it. Then ask for execution. This mirrors how strong coding agents behave in Plan mode.

7. Specify how to handle uncertainty

Add: "If you lack data, say UNKNOWN and list what you would need — do not guess."

8. Request self-check before delivery

End prompts with: "Before answering, list assumptions you made and flag anything unverified."

9. Use examples when format matters

One good example beats ten adjectives. Paste a sample email, SQL query, or landing section you want matched.

10. Set the verification step (human or automated)

Who checks? What proves success? A/B headline test, linter, legal review, or "compare output to source doc line by line."

11. Iterate with diffs, not restarts

Say "Change only paragraph 2 — shorten by 30% and keep the CTA." Do not re-run the whole prompt and lose good parts.

12. Know when to stop using AI

If the task needs real-time private data, binding legal advice, or lived domain judgment you do not have — AI assists prep, not the decision.

Pre-flight checklist

  • Job named with output format
  • Constraints listed
  • Source material attached
  • Plan approved (if high stakes)
  • Uncertainty rule included
  • Verification owner assigned

Copy-paste master prompt

Role: [operator type]
Task: [single deliverable]
Constraints: [bullets]
Sources: [paste below]
Output format: [table / steps / JSON / draft]
If uncertain: say UNKNOWN — do not invent.
Before final answer: list assumptions and open questions.

Common failure modes

SymptomFix
Generic fluffAdd audience + banned phrases
Wrong factsAttach sources; require citations
Overlong outputCap words; ask for outline first
Inconsistent tonePaste one approved example

Want more operator gear? Browse the MyGearHut free library or join The Gear Drop for weekly AI workflows — one email, no noise.

Want more gear like this?

The guide above is yours free — subscribe for weekly drops, templates, and tool updates.

One practical AI email a week. Unsubscribe anytime.