Engineering
Ground rules for building with Grok Bot + Cursor (it's an eager new developer)
Bot-written code works, but it duplicates tests, ignores existing structures, and can run up your cloud bill.
What hurt
Dave Wolf (@davewolfusa) shared what he has learned building with Grok Bot and Cursor: don't expect a well-architected codebase or reuse by default, so set ground rules. One night of de-duplicating cut his test cases from 790 to 328. Not affiliated with xAI.
What to install or create
Set rules up front: reuse existing tests instead of adding duplicates; have it write architecture, design, use-case, and rules docs for you to review, and regenerate them after heavy correction; spell out exactly how a new wrinkle differs from what exists; use development-priced infrastructure (spin up, run, and destroy big servers). Don't rely only on a QA bot — run tests locally, and probe errors that don't make sense.
Bots
No new bot. Follow the first-run steps.
Plugins
GitHub
Optional
Files
Architecture, design, use-case, and rules docs the bot keeps up to date for you to review.
Routines
None on the first run. Watch one manual pass first.
The prompt
Ground rules and lessons (author-quoted)
Quoted from @davewolfusa — as published. Do not invent a second system prompt.
a) Don't create a new test case if there is already an existing one (last night, we reduced the number of test cases from 790->328, and tonight we'll search for more duplicates (the duplicates drive up the length of time to validate a build, so can end up costing you greatly). b) Discuss architecture and design; ask it to produce architecture, design, use case, and rules docs for you to review. Ask Grok Bot to regenerate them whenever you've been correcting lots of issues or adding new functionality. c) Be very specific when asking for a wrinkle to an existing functionality. Specify exactly what's different and how it's similar. d) While in development, use development-priced infrastructure. In AWS, use Lambda, Fargate, and Cursor's ability to create, configure, and destroy servers to build datasets that only need periodic updates. This is better than using Fargate because you have to pay to keep the container definition around, while Cursor can stand up, run, and destroy a huge server for pennies. 5) Don't just depend on a QA bot to do all of the testing; run tests locally. 6) If you're seeing errors that aren't making sense, start probing.
First run
- 1Read https://x.com/davewolfusa/status/2107292021916602567 (credit @davewolfusa).
- 2Paste his ground rules (below) into your coding bot's instructions or skill.
- 3Ask it to write architecture, design, use-case, and rules docs, and read them before more code gets written.
- 4Have it search for duplicate tests and merge them before adding new ones.
- 5Run the test suite locally yourself after each big change, and ask it to regenerate the docs after a round of fixes.
Approval fence
Servers cost money. Have the bot destroy what it spins up, and keep an eye on infrastructure spend while developing.
What to tweak
Treat it like an excited new hire: great output, but it can't read between the lines — so write the lines.
Source
Dave Wolf @davewolfusa · Original post
Not affiliated with xAI. We did not write this prompt. Third-party template terms still apply.
Pro pack
Want the full setup, not just the idea?
Pro packs give you the job text, prompts, schedules and pitfalls for popular use-cases.