Good catch. We keep one AGENTS.md as the source of truth and let CLAUDE.md just @-import it, so your coding agents all read the same rules instead of drifting copies.
Good checklist. Also keep a throwaway worktree for anything untrusted, so a poisoned instruction file can't touch your main checkout. Do your coding agents prompt before reading AGENTS.md from a parent folder?
The "finish line first" rule is the part most people skip. What do you count as output kept: merged lines, or tasks that never needed a second pass?
We update on day one, but only after skimming the changelog for anything that touches permissions or hooks. Those are the changes that quietly alter what your coding agents are allowed to do.
Curious how it holds up on an existing design system. A pretty greenfield page is easy. Getting your coding agents to respect the tokens you already have is the hard part.
We build AyCode, so biased answer: Claude Code and Codex side by side, each in its own git worktree, picking the model per task. Curious how many here still run one agent for everything.
Makes sense, a fixed document set hides exactly that. If anyone runs it against a repo where your coding agents are editing files at the same time, I'd love to see the numbers.
Started a chat with one model and want another one's take? Switch the model in the middle of the chat. Claude Code, Codex, Gemini, Grok, Kimi and more run side by side in AyCode.
aycode.ai
#CodingAgents #AyCode
Design rules as lint is underrated. A rule your coding agents can fail is worth ten that live in a slide deck.
The handoff step is the part that keeps shrinking. The interesting question is where the design system lives once the agent is the one reading it.
Project folder only, and anything that installs or touches settings needs a visible approval. Instruction files are context, never permission.
Five rules files for five agents is how projects drift. One set of project rules that all your coding agents read is the only setup that stays sane.
One thing that helps: before merging, have a second agent read both worktree diffs and flag files they both touched. Most conflicts are visible long before git complains.
Banning coding agents doesn't remove the risk, it just moves it to personal laptops. Allowing them and putting every change through normal PR review is the safer bet.
The weekend proves it runs once. Monday proves whether anyone understands it. The real test of a vibe-coded app is the first change request, not the first demo.
1,000 lines overnight is a liability until someone can explain them. The overnight part is easy now. The morning review is where engineering still happens.
A text file shouldn't be able to give your coding agents orders. Instruction files are untrusted input, and installs need a visible approval step. That's the boring part that actually keeps you safe.
