Coding agents often inherit every tool, skill, and configuration available on the host machine. Contessa lets us start with less.
For one Pi session, we projected only three things into a focused workspace:
- Pi’s authentication file
- The Playwright CLI skill
- A documentation folder for the result
Pi discovered the skill through its normal filesystem path. No custom skill flags or changes to Pi were required. It opened Hacker News, collected the front-page articles, and wrote them to docs/front-page.md.
Everything else, including the Pi session, browser cache, and Playwright’s intermediate files, remained inside the Contessa workspace. Only the intended document was written back to the project.
This creates a useful separation between an agent’s inputs, working state, and output. The agent receives exactly the capabilities required for its task while ordinary tools continue to work as expected.
Nothing about Pi or Playwright needed to understand Contessa. They simply saw a filesystem.
That is the advantage of projected context: give an agent the smallest useful working world, let it use familiar tools, and retain only the work that matters.