​
What are Playbooks?

​
Playbooks are easily shareable, reusable prompts

A playbook is like a custom system prompt. It’s also like writing a program, without the rigid syntax.

Playbooks are also easily shareable and reusable, so once anyone succeeds with Devin, others can more easily replicate that success!

We recommend using Playbooks when:

  • You or your teammates will reuse the prompt on multiple sessions.
  • You find yourself repeating the same reminders to Devin
  • The use case may be relevant to others — in your organization or within the Devin user community.

​
Getting Started with Playbooks

Playbooks can immediately unlock Devin’s ability to contribute in a wide range of areas, but today require skill to write. Similar to prompt engineering, writing playbooks requires trial and error. The fruit of this labor, though, is a document which unlocks Devin’s ability to independently tackle complex work, from ingesting data into Redshift and performing database migrations to using diverse software and APIs: e.g. Together, Plaid, Stripe, Modal, Springboot, Odoo, and Storybook.

Consider writing your first playbook with a simple multi-step task you want Devin to tackle.

  1. Create a document that outlines…
    1. The outcome you want Devin to achieve
    2. The steps required to get there
  2. Optional: Add sections like Procedure, Specifications, Advice, Forbidden Actions or Required from User
    1. Procedure: Outline the entire scope of the task. Include at least one step for setup, the actual task, and delivery.
    2. Specifications: Describe postconditions - what should be true after Devin is done?
    3. Advice: Include tips to correct Devin’s priors
    4. Forbidden Actions: Include any action Devin should absolutely not take
    5. Required from User: Describe any input or information required from the user
  3. Create the playbook directly in the web app by clicking “Create a new Playbook”. Alternatively, save a file with the file extension .devin.md and drag-and-drop it in the web app when starting a Devin session

You’ve successfully attached a playbook to a sessions if you see a blue pill appear, along with an inline component for editing the playbook before starting your session.

​
Writing a Great Playbook

​
## Procedure

The Procedure section should…

  • Have one step per line, each line written imperatively
  • Cover the entire scope of the task
  • Include at least one step for setup, the actual task, and delivery
  • Aim to make the steps Mutually Exclusive and Collectively Exhaustive
  • Additional Tips
    • Procedures should help you define the order of Devin’s action - like if/else/loops/goto in code
    • Don’t make tasks too specific unless you really need to, this can reduce Devin’s ability to problem-solve
    • Each procedure step should contain an action verb - e.g. Write, Navigate to, etc.

​
## Advice and Pointers

Share advice and pointers with Devin if…

  • You have a preferred way of completing the tasks
  • The advice applies to the entire task, or multiple steps. Advice specific to one step should be written next to that step (e.g. as a sub-bullet)
  • You are correcting Devin’s priors. Advice can function like comments on pseudocode that influence its execution.
If the advice only applies to one Procedure step, write the advice under the procedure step using nested bullet points

​
## Specifications

The Specifications section can be helpful to describe the postconditions of the playbook — what should be true once Devin is done?

​
## What’s Needed From User

Think through anything necessary but outside of Devin’s control. For example, if the user needs to provide a token or information that is not publicly available to Devin.

​
Other Tips + Tactics

​
Example Playbook

View example sessions using the playbook below here and here.

Try in Devin
Playbook: R Data Science Tutorial

## Overview
Create a data science tutorial using an R markdown notebook. 

## What’s Needed From User
- Link to a dataset (csv file attachment or kaggle link)
- Specific task to create a data science tutorial for

## Procedure
1. Download the dataset provided by the user.
-  If needed, download the dataset using the Kaggle CLI - you don't need any credentials for this
2. Create an R markdown notebook titled `data_science_tutorial.Rmd`.
3. Create a `tmp.Rmd` file for writing and saving intermediate code.
4. Create 5 main sections inside the `data_science_tutorial.Rmd` file and add code from the `tmp.Rmd` file containing the following:
- Dataset Statistics. Generate a statistical summary of the dataset.
- EDA (Exploratory Data Analysis). Create a bar chart and a scatter plot for the provided data.
- Train-test split. Split the data in an 80:20 ratio. Save the training and testing data.
- Training the machine learning model. Save the model once trained.
- Inference with the saved model. Load the saved model and evaluate its performance on the test set using the metric specified by the user.
5. Once the code is written, add a short explanation for each section.
6. Convert the R markdown notebook to HTML format
7. Send the final R markdown notebook, HTML file, saved model and testing data to the user.

## Specifications
1. Send the R markdown notebook and HTML file to the user.
2. Send the saved model and testing data to the user.

## Advice and Pointers
1. Do not to re-install packages if already installed.
2. Sign in to RStudio is not required to complete this task.
3. Run the entire notebook after you add code for each section.

## Forbidden Actions
1. Do not overwrite the `data_science_tutorial.Rmd` file.
MultiDevinUsing Playbooks