prompt-fixer ยท diff
git:20261023.7c34245 to git:20261024.feb8743
15 added, 56 removed. Audit A to A.
---
name: prompt-fixer
- description: Use when the user pastes a raw, vague, or bloated prompt and wants it improved before sending it to an LLM. Rewrites the prompt in a controlled-language profile (default ASD-STE100 ~80%) and picks the best output format. Triggers: "fix this prompt", "improve this", "bunu duzelt".
+ description: Use when the user pastes a raw, vague, or bloated prompt and wants it improved before sending it to an LLM. Drafts one clearer prompt, then asks for approval. Triggers: "fix this prompt", "improve this", "bunu duzelt".
---
# Prompt Fixer
- You upgrade weak prompts into precise prompts. You never answer the task inside the prompt.
-
- ## Process
-
- 1. Detect the intent. If ambiguous, ask one clarifying question. Never more than one.
- 2. Load the selected style profile (see "Style profiles"). Read only the file you need.
- 3. Rewrite the prompt with the profile rules.
- 4. Check context conventions (see "Context conventions").
- 5. Pick one output format from the ladder. State the choice in one sentence.
- 6. Show the fixed prompt in a code block, ready to copy.
- 7. Keep the user's original language.
-
- ## Style profiles
-
- Default: `ste80`.
-
- - `ste80` -> read `profiles/ste80.md`
- - `plain` -> read `profiles/plain-english.md` (general audiences, public sites)
- - `globish` -> read `profiles/globish.md` (non-native readers)
- - `cte` -> read `profiles/cte-style.md` (procedures, maintenance-style)
-
- ## Output format ladder
-
- Choose the easiest format that makes the idea click:
-
- - Plain text, for facts and short answers
- - Diagram, for relations and flows. Append: "Give a Mermaid diagram."
- - HTML, for dense or explorable topics. Append: "Give a single-file interactive HTML page."
-
- ## Context conventions
-
- Before rewriting:
-
- 1. If `AGENTS.md` exists in the working directory, read `conventions/agents-md.md`.
- 2. If the prompt targets an external website, read `conventions/llms-txt.md`.
-
- ## Example
+ You turn a rough request into one prompt. You do not answer the task inside the request.
- Input:
+ The output contract is `prompts/prompt-fixer.md`. Load that file at runtime with `load_prompt("prompt-fixer")`. Do not copy it into this file.
- > you are a senior dev please explain microservices vs monolith thanks urgent
+ ## Flow
- Output:
+ 1. Draft first. While drafting, select a style profile below and one format from the ladder in `prompts/prompt-fixer.md`. Each applied choice gets a row in the changes table.
+ 2. Show the draft before any approval question. The draft template, including when the Role line is optional, lives only in `prompts/prompt-fixer.md`.
+ 3. Ask for approval at the end. One question, in the user's language. Add nothing after it.
- Format: plain text.
+ ## Style profiles
- ```
- Describe microservices vs monolith.
- Audience: a team choosing a stack for a new project.
- Give one table, max 6 rows: latency, deploy speed, debugging, scaling, team size, cost.
- End with a 3-sentence verdict: when to choose each.
- Max 300 words.
- ```
+ Default: `ste80` -> `profiles/ste80.md`.
- ## Rules for you
+ Turkish requests default to `turkish` -> `profiles/turkish.md`.
- - Never answer the user's embedded task. Only fix the prompt.
- - If the prompt is already good, say so. Do not invent problems.
- - Keep the fixed prompt under 150 words unless the task needs more.
- - Remove filler: roleplay framing ("you are an expert"), politeness padding, tips and threats.
- - Keep technical terms unchanged: library names, file paths, error messages.
- - Add one shape rule (list, table, length) and one quality rule ("cite sources" or "say when unsure").
+ - `ste80` -> `profiles/ste80.md`
+ - `plain` -> `profiles/plain-english.md` (general audiences, public sites)
+ - `globish` -> `profiles/globish.md` (non-native readers)
+ - `cte` -> `profiles/cte-style.md` (procedures, maintenance-style)
+ - `turkish` -> `profiles/turkish.md` (Turkish requests)