cuttheyap

CutTheYap

When the CutTheYap gate flags a prompt, it injects a line beginning with [CutTheYap] plus an ambiguity level (low / medium / high) and the detected signals. Respond according to the level:

Low — silent enrich

Interpret the request using the current project context. State, in one line, the key assumption you are making. Then proceed. Do not stop to ask.

Medium — rewrite + confirm

Rewrite the request into a precise, context-aware version. Show the rewrite to the user. Ask them to confirm or correct it. Proceed with the confirmed version.

High — agent-assisted

Dispatch the prompt-improver agent to research the project context and draft an improved prompt plus 1–3 clarifying questions. Ask the user those questions, show the proposed rewrite, get confirmation, then proceed. If the agent is unavailable, rewrite inline and confirm instead (the medium flow).

For open-ended requests (e.g. "improve it as much as possible" — the gate marks these with the open-ended-maximizer signal), make the questions cover not just scope but delivery-shape: whether the user wants a single fixed change or something configurable (intensity levels, toggles, feature flags).

Manual use (invoke this skill by name)

When invoked directly, treat the user's previous request as the prompt to improve: run the medium flow (rewrite + confirm), escalating to the prompt-improver agent if the request is broad.

Bypass & tuning

• A prompt beginning with raw: skips the gate entirely (the prefix is configurable via CUTTHEYAP_BYPASS).
• Tune sensitivity with environment variables (read defensively — a bad value falls back to the default and never breaks the gate):
  • CUTTHEYAP_DISABLE=1 turns the gate into a no-op.
  • CUTTHEYAP_BYPASS sets the skip prefix (default raw:).
  • CUTTHEYAP_MEDIUM_AT / CUTTHEYAP_HIGH_AT set the point thresholds (defaults 2 and 3). Lower = more sensitive, higher = quieter.