Free guide · Approving an AI plan before execution
Plan First: Prompt and Approval Checklist
Use this prompt when you want the AI to propose an approach before it produces the deliverable.
1. Define the job
Copyable template 1
JOB Outcome needed: Audience: Decision or action this supports: Inputs available: Constraints: Deadline: Required format: Known risks: Unknowns: What must not happen:
2. Ask for a plan, not the deliverable
Copyable template 1
Do not create the final deliverable yet. First propose a plan that includes: 1. your understanding of the outcome and audience, 2. the inputs you will use, 3. the steps you will take, 4. the proposed structure of the output, 5. the checks you will run, 6. any missing information or risky assumption, 7. anything that requires my approval. Keep the plan concise. Separate facts from assumptions. Wait for my approval before executing it. [PASTE JOB]
3. Review the plan
### Purpose
- □ The plan solves the requested job, not a nearby problem.
- □ The audience and decision are correct.
- □ The proposed output is usable in the required setting.
### Inputs and evidence
- □ Every required input is available or marked missing.
- □ Facts and assumptions are separated.
- □ Figures will be checked against a named source.
- □ No confidential data is requested without a clear need.
### Scope and sequence
- □ The steps are in a sensible order.
- □ The plan does not add unrequested work.
- □ High-risk or irreversible actions require approval.
- □ A human review happens before anything is sent, published or changed.
### Quality checks
- □ The plan names the checks it will run.
- □ Tone, length and formatting requirements are included.
- □ The plan explains what it will do when information is missing.
4. Approve with corrections
Copyable template 1
Revise the plan with these corrections: - - Return the updated plan only. Do not execute it yet.
When the plan is ready:
Copyable template 2
Plan approved. Execute only the approved plan. Stop and ask before any step that: - changes a live file or system, - sends or publishes content, - uses a new source, - expands the scope, or - requires a material assumption. Return the deliverable followed by a short completion report: steps completed, sources used, checks passed and open issues.
5. Worked example
Fictional job
Copyable template 1
Outcome needed: A recommendation on whether to extend the Northstar onboarding pilot. Audience: Operations director. Decision or action this supports: Approve or decline a two-week extension. Inputs available: Pilot tracker and team lead notes. Constraints: Use only verified pilot data; 1 page maximum. Deadline: Thursday at 4:00 p.m. Required format: Decision brief with recommendation, evidence, risks, owner and next date. Known risks: Completion counts may include duplicate tests. Unknowns: Time saved has not been measured. What must not happen: Do not claim cost savings or customer impact.
Useful plan
Copyable template 2
1. Confirm the tracker date and remove duplicate tests from the completion count. 2. Extract the invitation, completion and follow-up figures with source locations. 3. Summarize the team lead notes as observations, not measured outcomes. 4. Draft a one-page brief: decision, evidence, recommendation, risks, owner and review date. 5. Check every figure against the tracker and remove any savings claim. 6. Ask for approval before sending or updating any project file. Missing information: the intended owner of the extension and the date of the next review.
Why it works: it exposes the duplicate-count risk and two missing decision details before drafting.
Final check
- □ I approved the latest version of the plan.
- □ The final output follows that plan.
- □ Any deviation is explained.
- □ No live action happened without approval.
- □ I reviewed the final content in its real context.
The rule to keep: Approve the approach before you pay for the output in time, trust or rework.