PromptivityThe On-Task Kit: Rule files that keep AI coding agents on task
by Promptivity
$12
Buy on GumroadGet Scope Lock free
Short rule files that ask your AI coding agent to stay inside the task, ask before risky steps, stop when it is going in circles, prove it is done, and leave notes for the next session.
The On-Task Kit is a few short rules that ask your AI coding agent to stay on task, in the files your tool already reads.
Here's the problem it's built for. You ask for one small change. The agent edits five other files, says "done" when the test is still failing, or tries a sixth fix for the same error. Next session it has forgotten what you agreed.
Copy the files into your project (or paste the text into your app builder's settings) and type "guardrails check". The agent tells you which rules it loaded.
The five guardrails
- Scope Lock. Asks the agent to say what it will change before it starts, change only those files, and ask before going further.
- Session Handoff. Two small notes files that help a new session pick up where the last one stopped, and ask the agent to check with you before undoing settled choices.
- Loop Breaker. Asks the agent to stop and report after two failed fixes on the same error, instead of trying again.
- Done means done. Asks the agent to say "done" only after a real check passed, and to show the command and the result. It also asks for no weakened tests or placeholders.
- Ask first. Asks the agent to wait for a clear "yes" before deletes, migrations, force pushes, secrets, dependency changes, deploys or spending.
What's inside
- A ZIP of ready files: AGENTS.md, CLAUDE.md, Cursor rules (.cursor/rules/), .github/copilot-instructions.md, paste-ins for Lovable, Bolt and Replit, session notes templates, checklists and copy-paste prompts.
- A 9-page PDF: how to install, the five guardrails as checklists, a map of which tool reads which file and how to check it loaded, and what was tested live.
File names, locations and limits were checked against each tool's official docs on 19 October 2026. Cursor was tested live (5 of 5 behaviour tests passed, plus a re-run on the final rules). In the re-run, the agent warned before editing but didn't wait for a reply. Claude Code and Codex were checked to load the files; their behaviour was not tested. GitHub Copilot, Replit, Lovable and Bolt were checked against their docs only.
Rule files guide the agent. They are not locks. For anything that must never happen, also use your tool's own permission or approval settings.
Who it's for
- You build with an AI coding agent in Cursor, Claude Code, Codex or GitHub Copilot, or in Lovable, Bolt or Replit.
- You want the agent to stay on task without writing your own rules from scratch.
- You want a simple way to check that the rules loaded.
Not a coding course, and not tied to any language or framework.
Support
Questions or a broken download? support@promptivity.ai
Checkout and delivery are handled by Gumroad.
Not affiliated with or endorsed by any of these companies. Cursor is a trademark of Anysphere, Inc. Claude and Claude Code are trademarks of Anthropic, PBC. Codex is a trademark of OpenAI. GitHub and GitHub Copilot are trademarks of GitHub, Inc. Replit is a trademark of Replit, Inc. Lovable is a trademark of Lovable Labs Incorporated. Bolt is a trademark of StackBlitz, Inc.