---
title: "Prompt Improver"
---

# Project: Prompt Improver (The Orchestrator)

**Source:** `3-prompt/`  
**Model:** `gemma4:26b`  
**Type:** Bash script

## What It Does

This is the master controller — the script that ties everything together. It supports two modes:

1. **Regular mode:** runs a prompt through Ollama and saves the output
2. **Rerun mode:** combines the original prompt, test output, and generated code, then feeds everything back for improvement

After either mode, it runs the code extractor and any test file.

## The Prompt

```
Create a bash script with a shebang on the top that takes in a directory name

and optional a --rerun flag with a message

if its regular it runs

cat directory/prompt.md | ollama run gemma4:26b --nowordwrap | tee directory/ollama-output.md

if its a rerun it does

./1-combind/code \
    -m rerunmessage \
    directory/prompt.md directory/test-output.md directory/code | \
    ollama run gemma4:26b --nowordwrap | \
    tee directory/ollama-output.md

after either of those passes, it then runs

./2-extract/code directory

then it checks for a directory/test file, and if it's there it
chmod +x directory/test
directory/test 2>&1 | tee test-output.md

output only the code
```

## The Code

```bash
#!/bin/bash

DIR=$1
MODE="regular"
MESSAGE=""

if [ -z "$DIR" ]; then
    echo "Usage: $0 <directory> [--rerun <message>]"
    exit 1
fi

if [ "$2" == "--rerun" ]; then
    MODE="rerun"
    MESSAGE="$3"
fi

if [ "$MODE" == "rerun" ]; then
    ./1-combind/code \
        -m "$MESSAGE" \
        "$DIR/prompt.md" "$DIR/test-output.md" "$DIR/code" | \
        ollama run gemma4:26b --nowordwrap | \
        tee "$DIR/ollama-output.md"
else
    cat "$DIR/prompt.md" | ollama run gemma4:26b --nowordwrap | tee "$DIR/ollama-output.md"
fi

./2-extract/code "$DIR"

if [ -f "$DIR/test" ]; then
    chmod +x "$DIR/test"
    "$DIR/test" 2>&1 | tee test-output.md
fi
```

> **Note:** The original generated script had a typo in the rerun branch — `--nowordlamwrap` instead of `--nowordwrap`. This is why testing is essential! The code above has been corrected.

## Usage

### Regular Mode (First Run)

```bash
# Create a prompt directory
mkdir -p my-project
echo "Write a bash script that prints 'Hello World'" > my-project/prompt.md

# Run the prompt
./3-prompt/code my-project
```

This runs the prompt through Ollama, saves the output, extracts any code, and makes it executable.

### Rerun Mode (Improvement Loop)

```bash
# After the first run reveals bugs, feed back the error
./3-prompt/code my-project --rerun "The script has a syntax error on line 5: missing semicolon"
```

This packs the original prompt, test output, and code together, sends it to Ollama with your feedback, and produces an improved version.

## The Full Workflow

```
mkdir my-project
echo "prompt goes here" > my-project/prompt.md

# Step 1: Generate
./3-prompt/code my-project

# Step 2: Test (if you have a test file)
# The runner does this automatically

# Step 3: Evaluate
cat my-project/test-output.md

# Step 4: Rerun with feedback
./3-prompt/code my-project --rerun "the output formatting is wrong"

# Repeat steps 2-4 until it works
```

## Key Lessons

- **This is the orchestration layer.** It coordinates the combiner, Ollama, extractor, and test runner.
- **Two modes = two phases.** Regular mode for initial generation, rerun mode for iterative improvement.
- **Bash is enough.** The entire pipeline runs in bash — no Python, no frameworks, no build steps.
- **The rerun message is a human-in-the-loop channel.** The AI gets the files, but you provide the intent.

**Next: [Ollama REPL →](/run-ai-locally/projects/ollama-repl/)**