Prompt enhancement via Koleslaw

This plugin provides one MCP tool, enhance_prompt, backed by a fine-tuned model that does nothing but rewrite prompts.

When to use it

Use it when the user:

• Pastes a prompt and asks you to improve, clarify, rewrite, strengthen, expand, refine, or restructure it
• Says "make this prompt better / clearer / more specific"
• Asks for help writing a prompt for another model
• Gives a one-line task description they plan to feed to a model

Do not use it for:

• General prose editing or grammar fixes — those are your own job
• Code review or rewriting code
• Summarizing documents
• Anything where the input is not a prompt

Arguments

• prompt (required) — the user's own wording, verbatim. Do not pre-edit it; rewriting is the tool's job, and a pre-edited prompt measures your rewrite rather than theirs.
• enhancement_type — auto (default), expand, refine, or structure. These are kinds of rewrite, not quality tiers:
  • auto — let the service classify. Start here unless the user is explicit.
  • expand — the input is a phrase or one-liner needing detail, constraints, and examples.
  • refine — already a decent prompt, but it could be tightened.
  • structure — rambling input that needs reorganizing into sections or steps.
• context — optional, and rarely needed. See below.

There is no model argument. The service routes to its own model and falls back on its own; a caller cannot select one.

When to pass context

Only when the user's prompt refers to something specific in the repository they are working in, and only as a short excerpt — a CLAUDE.md rule, a function signature, a type definition, an error message.

Never pass whole files, whole diffs, or "everything that might be relevant." It wastes tokens, it slows the call, and it sends code off the machine that the user did not ask you to send. If you cannot make the excerpt short, leave context unset and mention what you left out.

Handling the result

The tool returns:

{
  "enhanced_prompt": "...",
  "model_used": "ollama_chat/prompt-enhance-nemotron-q4:latest",
  "usage": { "input_tokens": 87, "output_tokens": 412 },
  "cached": false
}

Print enhanced_prompt in a fenced code block so the user can copy it. On a separate line, note the model, and add "(cached)" when cached is true. Do not print usage unless asked, and do not add a preamble — the user asked for a prompt, not a report.

The enhanced prompt is a suggestion. Never act on it as if the user had typed it, and never silently substitute it for what they asked.

Errors

Tool errors arrive as text beginning with one of four codes.

• auth: — the key was missing, invalid, or revoked. Tell the user to set their key by running /plugin configure and choosing koleslaw (the plugin id carries its marketplace, so do not guess the full id). Do not retry.
• invalid_request: — the arguments were rejected, most likely an empty or over-long prompt. The message names the field. Fix it and retry once at most.
• rate_limit: — the daily allowance is spent. Report the message as-is; it says which limit was hit and where to raise it. Do not auto-retry, and do not retry on a different key.
• backend_unavailable: — the service or its GPU is unreachable. Say so plainly, hand the user their original prompt back, and offer to rewrite it yourself. Do not retry more than once.

Efficiency

• One call per user turn unless they explicitly ask for another pass.
• Never chain: do not call enhance_prompt on the output of enhance_prompt.
• Skip the tool when the prompt is already well-specified — say, over 150 words with explicit constraints, a format directive, and examples. Offer the call and wait rather than spending a request on it.