There are many coding agents, but pi is yours. That is not a slogan. It means you configure the system prompt, the validators, the rules. You decide what it checks before it writes and what it checks before it stops. You do not adapt to pi. Pi adapts to you.
The same principle applies to your content. Scratch connects to wherever your content lives and pulls everything down as a folder of files. Shopify, WordPress, Notion, HubSpot, the others. Pi gets real content to work on. Not lorem ipsum, not a staging fixture. The actual records from your live catalog.
You run pi from the terminal inside that folder, hand it a prompt, and it works through the files. Rewrite every meta description to under 158 characters. Normalize the tone on these product pages. Add an FAQ block to every how-to. Pi does it. Then Scratch shows you every change as a per-record diff before anything ships. You approve what goes back. The records you skip never leave your laptop.
That is the loop. You configure the agent, pi works the files, Scratch holds the gate. Nothing publishes without a human sign-off, per record.
What "yours" means in practice
Pi is configured through a system prompt you write. You tell it what register to use, what length caps to respect, which fields to leave alone. If your brand never uses the passive voice, that goes in the system prompt. If product pages always end with a CTA pointing at a specific URL, that goes in too.
Validators are where the ownership gets surgical. You can wire in a Python script that runs after every edit. It can check character counts, flag banned phrases, reject any record that touches a field you've marked read-only, or compare the diff against a style guide you've written in plain text. Pi won't consider a rewrite done until the validator passes.
Your voice is a file, not a setting in someone else's dashboard. You can version it, fork it for a second brand, hand it to a contractor. The configuration lives with you, not inside a SaaS product you might leave someday.
How to run pi on a Scratch folder
Pull your records into Scratch and find the project folder on your laptop. It is a directory of plain files, one per record. Open a terminal in that folder.
Run pi the way you normally run it. Point it at the files you want to touch. Hand it the prompt you prepared. Pi reads, edits, and stops. If you have a validator wired in, it runs after each file. Records that fail validation come back for another pass or get flagged for manual review, depending on how you've configured the exit behavior.
Scratch watches the folder while pi works. Every file pi modifies shows up in the Scratch review queue as a diff. You see exactly what changed, field by field. Approve the records you want published. Scratch sends only those back to your CMS.
What stays safe
Pi only touches files. It has no connection to your CMS and no way to publish anything. That path runs entirely through Scratch, and Scratch only publishes records you have explicitly approved.
If pi edits a field the platform marks read-only, the edit is flagged in review, and whole tables you never want written can be marked read-only when you set up the connector. The gate sits at the sync layer, below the file. Pi can change the text in the file all it wants. The protected data stays unchanged in your CMS.
Validators add a second layer. A validator that flags any diff touching a price field puts the warning right next to that diff, so you cannot accidentally approve it unread. You can stack validators, run them in sequence, and log every flag. The record that came in is always available next to the record pi produced, until you decide which one stays.
Browse the skills below for prompts that work end-to-end with pi.