translate

name	translate
description	Converts a prompt from one model's format to another. Re-selects techniques, adapts structure, preserves intent and domain content. Auto-triggers on: "/translate-prompt", "convert this prompt to GPT", "adapt for Claude", "port this prompt to Gemini", "make this work on o3", "translate prompt".
allowed-tools	Bash(python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/self-eval.py *) Read Write Edit

Prompt Translator

Convert a prompt optimized for one model into the optimal format for a different model. Preserve all intent and domain content — only change structure, techniques, and formatting.

Step 1: Load Source Prompt

Accept: file path, folder path, prompt name, or pasted prompt text.

If loading from a saved prompt folder, read metadata.json to get the source model and techniques.

Step 2: Identify Source and Target

• Source model: detected from metadata, format cues (XML → Claude, Markdown → GPT), or ask the user
• Target model: the user specifies which model to translate to. If not specified, ask.

Read both model entries from ${CLAUDE_PLUGIN_ROOT}/../../shared/models-registry.json.

Step 3: Diff the Models

Compare source and target across these dimensions:

Dimension	Example diff
Format	XML tags → Markdown headers (Claude → GPT)
Reasoning	Extended thinking → standard (Opus → Sonnet: remove "think thoroughly", add explicit CoT)
Reasoning	Standard → reasoning-native (GPT → o3: remove ALL CoT, strip examples)
Few-shot	Required → avoid (Gemini → o3: remove all examples)
Few-shot	Not required → required (Claude → Gemini: add examples if missing)
Tone	Neutral → explicit (Claude → GPT-4.1: add sandwich method, repeat key constraints)
Length	Any → minimal (anything → o-series: strip to essentials, <200 words)
Special	Any → image-gen: restructure as descriptors, remove all instruction scaffolding

Step 4: Apply Translation

Read reference files for the target model:

• model-profiles.md
• technique-engine.md

Apply these transformations in order:

Format Conversion

From	To	Action
XML (<instructions>, <context>)	Markdown	Convert tags to # Role, ## Instructions, ## Examples headers
Markdown (# Role, ## Task)	XML	Wrap sections in <instructions>, <context>, <constraints>, <examples> tags
Any structured format	Minimal (o-series)	Strip all markup. Keep only the core instruction in plain text. Remove examples, CoT, and verbose constraints.
Any format	Descriptors (image-gen)	Extract visual description, style, and constraints as comma-separated descriptors

Technique Re-selection

Source technique	Target model type	Action
Chain-of-Thought	reasoning-native (o3, R1, QwQ)	REMOVE — built-in reasoning, explicit CoT hurts
Chain-of-Thought	extended-thinking (Opus)	REPLACE with "think thoroughly"
No CoT	standard (GPT, Sonnet)	ADD "Think step by step" for complex tasks
Few-Shot examples	o-series	REMOVE — zero-shot outperforms
No examples	Gemini	ADD 2-3 examples — required by Google guidance

Model-Specific Adaptations

Target	Adaptation
GPT-4.1	Sandwich method: repeat critical constraints at the end
Claude	"Think thoroughly" not "step by step". Avoid MUST/ALWAYS. Use XML tags.
o-series	Developer messages, not system. "Formatting re-enabled" if markdown needed. <200 words.
Gemini	Temperature 1.0 note. Always include examples.
DeepSeek R1	Treat like o-series. No explicit CoT.

Step 5: Validate Translation

Run self-eval on the translated prompt:

python ${CLAUDE_PLUGIN_ROOT}/../../shared/scripts/self-eval.py <translated-prompt>

If the translated prompt scores LOWER than the source, warn the user and explain which dimensions degraded (e.g., "Model Fit improved 5→9 but Completeness dropped 10→7 because examples were removed for o-series").

Step 6: Save

Save the translated prompt as a new folder: prompts/<name>-<target-model>/

Run the full delivery pipeline (token count, metadata, tests, report).

Present both prompts side-by-side in a comparison:

TRANSLATION: stocks-analysis (Claude Opus → GPT-4.1)

  Source: prompt.xml (Claude, XML, 1326 tokens, 9.3/10)
  Target: prompt.md (GPT-4.1, Markdown, 1280 tokens, 8.9/10)

  Changes applied:
    ✓ XML tags → Markdown headers
    ✓ "Think thoroughly" → "Think step by step"
    ✓ Added sandwich method (constraints repeated at end)
    ✓ Added explicit output format section
    ✗ Slight Completeness drop — GPT-4.1 needs more literal instructions

  Recommendation: run /converge on the translated prompt to push score higher.

Rules

• NEVER change the prompt's intent, domain content, examples, or task description.
• Only change structure, format, technique scaffolding, and model-specific adaptations.
• If the source prompt has custom examples, translate the FORMAT of examples but keep the CONTENT.
• Always explain what changed and why in the comparison output.
• If translating to a fundamentally different domain (text → image-gen), warn the user that significant content loss is expected.