Simple technique,
proven results.

When auto_apply: true is set in your settings file, a hook intercepts every prompt you type before Claude processes it. The hook repeats your prompt based on your configured mode (simple, verbose, or triple), then Claude sees the repeated version.

Example: You type "What's the 5th item?" → Hook transforms it to "What's the 5th item?What's the 5th item?" → Claude processes the repeated version → You get improved accuracy.

The repetition is completely transparent - you just type normally and get better results automatically.

1. Copy the example settings file:

cp .claude/prompt-repeater.local.md.example .claude/prompt-repeater.local.md

2. Edit the file and change auto_apply: false to auto_apply: true

3. Restart Claude Code

Now every prompt you enter will automatically be repeated. To turn it off, change back to auto_apply: false and restart.

Best for non-reasoning tasks:

• Multiple choice questions
• List navigation (find Nth item, find between items)
• Simple fact retrieval
• Pattern matching

Less effective for:

• Multi-step planning
• Complex debugging
• Tasks already using chain-of-thought

The good news: It's safe even for reasoning tasks (neutral to slightly positive), so you can leave automatic mode on and it won't hurt performance.

Simple (recommended): <QUERY><QUERY> - Default mode, works well for most tasks

Verbose: <QUERY> Let me repeat that: <QUERY> - Sometimes marginally better with framing text

Triple: Three repetitions - Maximum accuracy for list navigation tasks, but longer input

Start with simple mode. Research shows all three perform similarly on most tasks, with triple providing the best results on specific list-heavy tasks.

It depends. Prompt repetition only affects the parallelizable prefill stage. Research shows no latency penalty for most models and prompt lengths.

The only exception: Claude models with very long prompts (>2000 characters) or triple repetition may see slight prefill latency increase. But generation speed is completely unaffected.

Yes! Keep auto_apply: false in your settings (or don't create a settings file at all). Then use these commands:

/repeat-last - Apply simple repetition to your last prompt
/repeat-verbose - Apply verbose repetition
/repeat-3x - Apply triple repetition

This gives you full control over when repetition applies. Great for experimenting to understand when it helps.

Prompt repetition duplicates your input so each token can attend to every other token, fixing positional limitations in the attention mechanism.

LLMs process tokens left-to-right where past tokens can't attend to future tokens. This means token order affects accuracy.

For example, "<CONTEXT> <QUESTION>" performs differently from "<QUESTION> <CONTEXT>" because tokens at the beginning can't see tokens at the end.

Prompt repetition solves this by duplicating the entire prompt, enabling each token to attend to every other token in at least one of the repetitions.

See the Google Research paper for the full technical explanation.

This is an early beta. The plugin implements a technique from a peer-reviewed Google Research paper published in December 2025.

Why we think it's safe:

• Based on research tested on 7 major models (Gemini, GPT-4o, Claude, DeepSeek) across 7 benchmarks
• 47 wins, 0 losses on non-reasoning tasks - never performs worse than baseline
• Safe even with reasoning tasks (neutral to slightly positive)
• Plugin uses safe bash scripting, validates inputs
• Doesn't modify any files or make network requests
• Only manipulates prompt text before sending to Claude

None, really. Prompt repetition is a simple technique with proven benefits and no downsides.

The only considerations:

• Input length doubles (or triples with triple mode), so very long prompts may approach context limits
• Effectiveness varies by task type (best for non-reasoning, neutral for reasoning)
• Requires restart to change settings (hooks load at session start)

But there's no performance penalty, no cost increase (beyond doubled input tokens), and never worse accuracy than baseline.