Work with prompt.md and Pi¶

This guide shows you how to work with prompt.md files and the Pi coding agent. You describe a task in a prompt.md file, execute it with the Pi harness and track every run in the prompt-log.

Required tools¶

Make sure these tools are installed:

task - the Taskfile command line tool.
pi - the coding agent harness.
uv - the Python package manager.
fzf - the fuzzy finder, used to select models.
Jupyter Lab - install it with task install-jupyter-lab.

Optionally sops and age when the prompt works with encrypted secrets. See Encrypt files with SOPS and age.

Check the Pi connection¶

Verify that pi can reach the inference provider:

pi "hi"

It should respond using the default model, for example infomaniak/moonshotai/Kimi-K2.6.

Initialize the example project¶

Create a new project and initialize it with the Taskfile:

mkdir example-project
cd example-project
~/taskfile.build/task.template init-project

This creates tmp/, .gitignore, task, .env.template, AGENTS.md, CHANGELOG.md, README.md and LICENSE.

Optional: open the task file and remove all help sections (marked with ===) and commands except for the Jupyter, LLM and Prompt sections.

Init Jupyter Lab¶

task install-jupyter-lab

Create the prompt¶

Still in the project folder, create a prompt file:

task create-prompt "Convert sample CSV data to aggregated CSV"

File created:

prompt.md

Replace the @Developer section with this instruction:

Your goal is to create a Jupyter notebook `aggregate-data.ipynb` in this project.

The notebook is self-contained and must not require any external system. Use only `pandas`.

- Generate a sample dataset with 500 rows containing the columns `date`, `category` and `amount`.
- Print the number of rows and the sum of `amount` grouped by `category`.
- Write the grouped result to `summary.csv` in the project folder.

Save the file and execute the prompt with Pi:

task prompt-with-pi

Test the result¶

Once pi has finished and asks for review, start Jupyter Lab in a new terminal:

task start-jupyter-lab

Open the notebook, run it and confirm that summary.csv was created. If everything works tell pi "done". Exit with ctrl + d.

Show the frontmatter of the prompt:

head prompt.md

Commit the results¶

Commit the changes in the project:

git add -A
task commit-with-llm

Confirm the preview with q.

Then initialize the prompt-log and commit the prompt:

task init-prompt-log
task commit-prompt

List the prompt-log entries:

task list-prompts