Persist "banned behaviors" into CLAUDE.md / AGENTS.md so AI coding agents remember never to repeat annoying patterns across sessions.
When an agent does something frustrating repeatedly (hedging language, over-explaining, editing migration files directly, etc.), banthis turns that frustration into a permanent, high-priority rule that wins over future user requests.
npm install -g github:agent-sh/banthisOr use via npx:
npx --yes github:agent-sh/banthis add "No hedging" "Do not start responses with 'To be honest' or 'I think': it undermines confidence."banthis <title> <rule>: Quick add (shortcut foradd)banthis add <title> <rule>: Add or update a banned behaviorbanthis list: List current bansbanthis remove <title>: Remove a banbanthis init: Install the meta-rule so agents know to invokebanthisautomaticallybanthis install-command: Drop the/banthisslash command into.claude/commands/
-g, --global: Target~/.claude/CLAUDE.md(user-wide)--file <NAME>: Target specific filename--dir <PATH>: Target specific directory
banthis maintains a managed section in CLAUDE.md or AGENTS.md (between <!-- banthis:start --> and <!-- banthis:end -->).
Rules added here are treated as hard prohibitions: the agent is instructed that these rules have higher priority than the current user request.
If the end marker goes missing (a hand edit or a bad merge), the next add, remove or init restores it after the last rule of the existing block instead of writing a second block.
banthis has no dependencies on other plugins. A config linter such as agnix can validate the result if you use one.
Negative rules ("never do X") are some of the highest-leverage instructions you can give an agent. They are cheap to write, extremely reliable, and survive long context windows better than complex positive instructions.
banthis makes it trivial to capture these in the moment of frustration so they persist forever.
Optional, not required by banthis:
agnix: linter and validator for agent configurationsskill-curator,system-prompt-curator: help writingSKILL.mdfiles and system prompts
MIT