A04: Context Engineering

The assistant is brilliant and has amnesia about your specific situation. It knows general things from training, but it does not know your file, your error, your folder, unless you show it. Putting the right information in front of it is called context engineering, and it is the difference between a guess and an answer.

Show, Don't Describe

Describing a file to the AI is like reading a document aloud over the phone: slow and lossy. Instead, hand it the file. In Antigravity CLI you do this with @:

Explain what @notes.txt is about in two sentences.

The @notes.txt pulls that file's actual contents into your message. Point at the real thing, @report.md, @script.js, or a whole folder with @src/, and the AI reads it directly instead of relying on your summary. Because Antigravity CLI runs in a folder, it can also see the files in the folder you started it from.

Use Your Real Files

Toy files are for practice; the point is your real work, an essay, meeting notes, a spreadsheet. On Mac and Linux they are already reachable. On Windows with WSL, your documents live under /mnt/c/, so C:\Users\Alex\Documents becomes /mnt/c/Users/Alex/Documents. cd into that folder (put quotes around any path with spaces) and @ a file like any other.

Plain text drops straight in, @notes.txt, @data.csv, @essay.md. A Word document or PDF is not plain text, so give the AI a text copy: in Word or Google Docs use File then Save As (or Download) as plain text or Markdown, then @ that copy. For PDFs, or many files at once, a converter like markitdown does it in one command (pip install 'markitdown[all]', then markitdown essay.docx -o essay.md). Either way you end up with clean text the AI reads perfectly and a copy you can edit.

Focused Beats Everything

More context is not better. There is a sweet spot:

• Too little - you paste an error with no code, and it guesses.
• Too much - you dump twenty files, and it drowns, gets slower, mixes unrelated things together, and eats through your daily free-tier limit faster.

Give what is relevant to this question and no more. When you paste an error, paste it exactly, word for word, not "it said something about a module." The exact text is the clue.

Start Fresh When You Switch Topics

A long conversation carries everything you said earlier as context. Helpful when it is related; confusing when you jump to a new topic and old details bleed in. When you change subject, start a new conversation so the AI is not still thinking about the last one.

This Week's Exercise

1. Create a short text file (some notes, a recipe, anything) in the folder where you run agy.
2. Ask a question about it without @, describing it in words. Note the answer.
3. Ask the same question with @yourfile. Compare, which answer actually knew what was in the file?
4. Now paste a real error message you hit this week (exact text) and ask what it means. Bring the comparison to class.