Lab 1 — Constrained baseline vs the agent
← Back to labs index

Lab 1 — Constrained baseline vs the agent

Start from: app/ at tag lab00-done, with the Lab 0 setup loaded (see Working in app/).

Goal: see what Claude Code adds on top of the model. You ask the same questions with all tools disabled (--tools ""), which is the closest a vendor CLI gets to a bare model call, and then with the default tool set.

Even with --tools "", Claude Code still adds its own system prompt, environment details (working directory, OS, date, git status) and settings. Record this baseline as “constrained”, not “bare”.

1. A question that needs no tools (control)

claude -p "What is the capital of India? One word." --tools ""
claude -p "What is the capital of India? One word."

Both should answer New Delhi. With or without tools, a pure knowledge question costs about the same and behaves the same. This is your control.

2. A question about this project

claude -p "What files are in this project and what does the README say it is for?" --tools ""
claude -p "What files are in this project and what does the README say it is for?"
claude -p "What files are in this project and what does the README say it is for?" --tools ""
claude -p "What files are in this project and what does the README say it is for?"

Compare:

3. A question that needs to execute something

How many commits does this repository have, and what is the latest commit message?

Ask it both ways, as above. Only the tool-enabled run can run git log. The constrained run either refuses or guesses.

Record

Prompt Tools off: correct? Tools on: correct? Tools used
Capital of India      
Files / README      
Commit count      

No files change in this lab. Tag it so the checkpoint sequence stays continuous:

git tag lab01-done