Who it is forA developer who keeps re-prompting because the agent declares success without proof.
The problemWeak criteria like make it work need constant clarification, and the agent stops before the job is really done.
The outcomeTasks phrased as verifiable goals, with a short plan where each step has a check, so the agent can loop on its own until they pass.
What you need first
- Claude Code (plugin install) or any project where the agent reads CLAUDE.md
- curl for the per-project install
- Cursor users can use the committed project rule in .cursor/rules
Steps
- Install the guidelines with the plugin or the CLAUDE.md option
- Rewrite your request as a success criterion, for example write a test that reproduces the bug, then make it pass
- Ask for a brief plan where each step ends with verify: and a check
- Let the agent loop until the checks pass instead of approving each step by hand
- Review the final diff against the success criteria you set
Prompt for your agent
Using the Goal-Driven Execution principle from andrej-karpathy-skills, turn this task into verifiable success criteria: <task>. Write the failing test first, give me a numbered plan where each step has a verify check, then loop until all checks pass.
Caveat.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for multica-ai/andrej-karpathy-skills and "Make your agent work toward tests instead of vague tasks"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer