Skip to content

Instantly share code, notes, and snippets.

Show Gist options
  • Select an option

  • Save ERICJ3ffrey/39a837602e804f177d3bfed26985840e to your computer and use it in GitHub Desktop.

Select an option

Save ERICJ3ffrey/39a837602e804f177d3bfed26985840e to your computer and use it in GitHub Desktop.
6 Prompts To Get Consistent AI Results
1. Parallel agents
Use parallel agents for this.
Split the work into four separate agents:
- one to understand the current setup
- one to find the simplest path
- one to look for risks and edge cases
- one to suggest how we verify the work
Keep their work separate, then summarize the best plan.
2. Spec first
Write the implementation spec before changing files.
Explain the goal, what currently happens, what should happen instead, what files may change, the risky parts, how we will verify it, and what
“done” means.
Wait for my approval before implementation.
3. Interview before build
Interview me before you build.
Ask only the questions needed to remove ambiguity.
Find out who this is for, what it should do, what already exists, what should change, what could go wrong, and what not to touch.
After my answers, summarize the implementation spec.
4. Verification plan
Before building, show me exactly how you will verify this.
Tell me what you will check, what result you expect, what failure looks like, and what needs human review.
Do not implement until the verification plan is clear.
### 5. Save project instructions
Based on this session, update the project instructions.
Save only durable rules: commands that worked, test commands, project conventions, gotchas, files not to touch, review checklist, and what done means.
Do not save temporary task details.
### 6. Automate or augment
Evaluate whether this should be automated or augmented.
Ask if this requires taste or judgment, if 80 percent quality is acceptable, what happens if the output is wrong, what should stay
human-approved, what logs or alerts are needed, and how we verify it keeps working.
Recommend automate, augment, or leave manual.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment