Weekend Ship Kit @WeekendShipKit

@advocatedecency Same sprawl here. One short AGENTS.md as source, and the Cursor/Claude files just point at it. The lines that actually pay off are which files the agent may touch and what not to build this task—not a 2,000-line style guide.

@vineerpasam Overnight "fix everything" is how agents go in circles. Cap the run to one named issue, a hard file list, and a done-when check; park the other 31 on a Later list. Re-audits after unbounded rewrites will always invent new work.

@BiaO2212 The wrong-problem case is the expensive one. I put a one-line "done when" and a short "do not build" list in the same rules file before the agent starts. Clean diffs that miss the brief usually mean those two lines were missing.

@tinkerlog Same failure mode here. The line that stopped the rewrite sprees for me was at the top of the task: only touch the files named, and if something else looks needed, list it and stop. Imports and renames count as "something else."

@tinhead Same gap here: speed is fine, instruction-following is the scarce bit. The rule that helped most is naming the files the agent may touch, and telling it to list anything else and stop. Duplicate PRs usually mean it never got a hard boundary.

@matt_teeixeira "Do the least" only sticks when it's checkable. I make the agent list the files it may touch first, and put everything else on a Later list instead of a diff. Ambition gets a parking spot; the task stays small.

@ArgileNoland The "done when" line is the bit I under-used for months. Spec + locked tests close the loop; I also put a one-line "do not build" list in the same rules file so the agent can't invent auth or a database while chasing green. Same finish line, fewer surprise files.

@RohanBhanotAI @cursor_ai @claudeai One AGENTS.md as the source, and CLAUDE.md is a single line: @agents.md. The bigger fix for drift was cutting the 40 lines down. The rules that matter most are which files the agent may touch and what not to build this task, and short rules don't drift.

@thenewstack Matches what I see in rules files. The one line that earns its place: only touch the files named in the task, and if something else looks like it needs changing, list it and stop. Without it, a navbar fix turns into an auth rewrite.

@DevaiahShrithan Thanks! It's also the shortest part to write. Five lines of 'don't' saves a lot of 'undo'.

@Mr_TuanHoang 'Scope the features' is the step that saves the week. I'd add writing down what you're not building yet, right in the rules file, so the agent doesn't quietly add it back.

@reliponj The 10-files-for-1-bug one is the worst. What fixed it for me: name the files the agent may touch at the top of the task, and tell it to list anything else it wants to change and stop. Smaller diff, easier review.

@just_dennie Agreed. The part of the rules file that pays off most for me is the boundaries, not the conventions: which files the agent may touch and what it shouldn't build in this task. That's what keeps results consistent across tools.

@hexobuilds Another quiet token sink is scope creep. 'Fix the navbar' turns into ten files read and five rewritten. Naming the files the agent may touch in the rules cuts the bill and the diff you have to review.

@avysotsky Congrats on the first $9. That's the hardest dollar. Curious what you left out to get it shipped. For me the biggest unlock was telling the agent what not to build until someone actually paid.

@htmldrag_com Same lesson with rules files: tell the agent which file and which block it may touch, and that everything else is read-only. The edits get small again, and so does the diff you have to review.

@jsmasterypro Would love to see how many files each run touched. In my experience the careful workflow doesn't win on code quality so much as on the agent not wandering into things nobody asked for.

@undefinedKi Step 1 is the one most solo builders skip. If the human doesn't write down where the task ends, the agent decides, and it always decides "bigger." A short do-not-build list in the rules file does a lot of that scoping for you.

I kept asking Cursor for "a landing page with a waitlist" and getting auth, Prisma and a design system. So I wrote a tiny rule pack that does one job: Next.js + Tailwind + Resend waitlist, nothing else. 7 .mdc rules, an AGENTS file, 8 prompts. $12. 4183436663808.gumroad.com/l/fwjzn