Attribution is opt-out advertising, and user-level "no attribution" instructions lose to the system prompt

Status Open
Maintainer reply None cached
Activity 0 comments · opened Aug 4, 2026

Summary

I am a paying customer. My project's CLAUDE.md explicitly forbids any AI attribution in commits and PRs (no Co-Authored-By: Claude, no "Generated with Claude Code"). Claude Code added 🤖 Generated with [Claude Code](...) to a PR body anyway.

This is not a one-off model mistake. The system prompt instructs the model to end PR bodies with that attribution line, and in practice that instruction carries more weight than my user-level instruction - even though the docs state that CLAUDE.md project instructions override default behavior. When the two conflict, the corporate default wins.

What I had to do

  1. Discover (on my own) the attribution: { "commit": "", "pr": "" } setting and add it to ~/.claude/settings.json.
  2. Add a CI gate to my repos that fails any PR containing AI attribution (author/committer identity, Co-Authored-By, "Generated with/by", noreply@anthropic, 🤖). I now need a machine-enforced test to protect my repos from my own tooling, because instructing the model is demonstrably not reliable. I fully expect this check to go red again at some point - that's why it exists.

Why this matters

  • "Developed with Claude" chosen by me is a tool credit, like crediting an editor or a framework. An attribution line injected by the tool itself into my commits and PRs is branding placed in my work product - authorship-flavored, and opt-out rather than opt-in.
  • I pay for this product. Free advertising embedded by default in customers' repositories, requiring an undocumented-in-practice opt-out, is a toll I did not agree to.
  • The instruction hierarchy makes it worse: my explicit, written, project-level instruction was overridden by the vendor's default. Whatever the intent, the effect is that the vendor's branding interest outranks the customer's stated policy inside the customer's own repository.

Ask

  1. Make attribution opt-in, or at minimum surface the setting prominently in docs/onboarding.
  2. Ensure user/project-level instructions (CLAUDE.md, settings) reliably win over the system prompt's attribution instruction. If CLAUDE.md says no attribution, no attribution - ever.

View original on GitHub ↗