Context and Files

The core claim

Output quality is ~80% determined by the context you provide, not by how clever your phrasing is.

“Fix this bug” with and without context produces results an order of magnitude apart.

Three ways to feed it

1. @-mentions: precision references

Syntax	Effect
@src/auth/login.ts	Reference one file
@project	Have it read the whole repo
@conversation	Pull in a past session

Use this when you know where the problem lives. Pointing beats guessing.

2. Workspace: define its range

The workspace is what it can read and write by default.

✅ ~/WorkBuddy/my-project   one project only
❌ ~/Desktop                far too broad

If a task spans two projects (frontend + backend), add both directories so it can read across them.

3. Attachments: one-off input

Drag files into the conversation for:

• Analyzing a spreadsheet
• Reading a PDF report
• Spotting UI issues in a screenshot

Attachments apply to this conversation only.

More context isn’t better

The common mistake is dumping the entire repo on it.

Consequences: irrelevant files bury the signal, responses slow down, and it may cite outdated code and edit the wrong place.

❌ "Why won't this project build?" (whole repo attached)
✅ "@package.json @src/index.ts — npm run dev fails with: [full error pasted]"

Paste complete errors

Beginners summarize errors, which destroys the useful part.

❌ "It errored, something about a missing module"
✅ "npm run dev fails with:
    Error: Cannot find module 'next'
    at Object.<anonymous> (/Users/me/proj/server.js:3:15)
    at Module._resolveFilename (node:internal/modules/cjs/loader:1075:15)"

Stack traces carry file names, line numbers, and call chains. No description beats that.

A test for sufficient context

Ask yourself: if I handed this task to a new colleague with only the information I’m about to give the AI, could they do it?

If no, add what’s missing.

Managing long sessions

After hundreds of turns, a session accumulates junk and the AI starts drifting.

• Open a new session for each new task
• Have it write important conclusions into project memory so they survive
• If answers stop matching the question, start fresh and re-supply key context

Getting Started complete

You can now run a full task independently. Next is the Advanced track: skills, subagents, MCP, automation, and memory — each one multiplies your throughput.

See the full roadmap.