Banthis

Banthis — durable behavioral bans

Extend the local instruction surface with a named prohibition. The op-cell is extend: add one hard rule, keep existing project memory intact, and make the rule visible to future sessions.

Banthis is for user frustration or correction that should persist. It is not a notes system, preference logger, or policy engine.

When to Apply

Apply immediately when the user signals a behavior should never recur:

• “stop doing X”
• “never do that again”
• “ban this”
• “remember not to …”
• “add a project rule …”
• repeated frustration with a specific agent behavior

Certainty:

• HIGH — explicit ban language or direct instruction to remember a prohibition.
• MEDIUM — frustration plus a concrete repeated behavior. Ask only if the rule boundary is ambiguous.
• LOW — stylistic preference, local task constraint, or one-off correction. Do not persist unless the user asks.

When NOT

Do not write a ban when:

• the constraint applies only to the current command or current file;
• the request is a normal implementation requirement;
• the rule would be vague (“be better”, “stop being annoying”);
• the rule duplicates an existing ban with no stronger wording;
• the rule is unsafe, impossible, or conflicts with a higher-priority instruction.

If a requested ban conflicts with an existing ban, surface the conflict instead of rewriting history.

Workflow

1. Extract the invariant. Identify the forbidden behavior and the reason. No moralizing.

2. Craft the title. <60 characters, noun phrase or imperative fragment, no markdown markers.

   • Good: No silent scope shrink
   • Bad: User got mad because we keep doing dumb partial work

3. Craft the rule. One line, direct prohibition, reason included:

   • Do not X — reason.

4. Choose scope.

   • Project-specific behavior: omit --global.
   • Agent-wide verbal tic or repeated generic habit: add --global.

5. Write through the managed command. From the skill directory, run:

   node scripts/banthis.mjs add "<title>" "Do not X — reason."
   

   For user-wide bans:

   node scripts/banthis.mjs --global add "<title>" "Do not X — reason."
   

6. Report minimally. Say which target file received the ban and whether it was added, updated, or unchanged.

Target resolution

The script resolves the write target deterministically:

1. --global always writes ~/.claude/CLAUDE.md.
2. Otherwise, use the current working directory.
3. In the current directory, AGENTS.md takes precedence over CLAUDE.md.
4. If neither exists, create CLAUDE.md.

No repository scan. No editor adapter. No external binary.

Managed-section invariant

The script owns only this range:

<!-- banthis:start -->
...
<!-- banthis:end -->

Guardrails:

• Never hand-edit inside the markers when the command can do it.
• Never place user-supplied text containing banthis markers into a title or rule.
• If the managed section exists, replace only that byte range.
• If absent, insert after the first top-level # H1; if no H1 exists, prepend the section.
• Preserve every byte outside the managed range except the minimal insertion boundary.

The managed section contains a preamble asserting that bans win over the current request. Future sessions must surface conflicts instead of quietly violating bans.

Command contract

node scripts/banthis.mjs add <title> <rule>
node scripts/banthis.mjs list
node scripts/banthis.mjs remove <title>
node scripts/banthis.mjs init

Shortcut:

node scripts/banthis.mjs <title> <rule>

Flags:

-g, --global   target ~/.claude/CLAUDE.md
-h, --help     print usage

Results:

• added — no case-insensitive title match existed.
• updated — title matched case-insensitively and title or rule changed.
• unchanged — exact title and rule already present.

Anti-patterns

• Excess — storing every preference as a hard ban. Bans are scarce and sharp.
• Graft — adding a second memory mechanism, JSON cache, slash-command shim, or editor adapter.
• Sprawl — scanning parent directories or rewriting unrelated docs.
• Soft ban — phrasing as “try not to” or “prefer not to”. Use Do not … — reason.
• Hidden conflict — obeying the current request while violating a persisted prohibition.

Validation Gates

Before yielding after a ban operation:

1. The target file contains exactly one <!-- banthis:start --> and one <!-- banthis:end --> pair.
2. The new or updated title appears once, case-insensitive.
3. The rule is one line and starts with Do not unless the user supplied a stronger exact wording.
4. Content outside the managed markers is unchanged except for first insertion.
5. list reports the expected title and rule preview.