Banthis

Banthis Skill

When the user shows clear frustration with a recurring agent behavior, you must capture it using banthis so the rule persists across all future sessions.

Parse Arguments

Use $ARGUMENTS as the behavior to ban. If it is empty, infer the behavior from the most recent explicit correction in the conversation.

$ARGUMENTS

When to Activate

• User says variations of: "stop doing X", "you always do X", "don't ever X again", "ban this", "remember not to..."
• User corrects the same class of mistake multiple times in one session
• You notice a pattern of low-value behavior (excessive hedging, over-explaining, touching files you were told not to, etc.)

How to Use

1. Do not ask for permission. The user has already signaled they want this behavior banned.
2. Craft two strings:
   • title: Short (under 60 chars), framed as a prohibition (e.g. No 'let me be honest' preambles)
   • rule: 1–2 sentences in the form Do not X — reason.
3. Call the tool (prefer the local banthis if available, otherwise npx --yes github:agent-sh/banthis):

banthis add "<title>" "<rule>"

Use --global only for behaviors that should apply to every project (verbal tics, generic LLM habits).

4. Confirm in one short line: Banned: <title>

Important Notes

• Rules added by banthis go into a managed section of CLAUDE.md or AGENTS.md and have higher priority than normal instructions.
• This is the negative counterpart to positive memory tools (e.g. axiom).
• Run banthis init in projects if the meta-rule is not yet present (it teaches agents to invoke banthis automatically).

Related Tools

• axiom (positive knowledge capture)
• agnix (validates the resulting CLAUDE.md / AGENTS.md)
• skill-curator and system-prompt-curator (for improving how agents are instructed)