A trick from my friend Paul Tobias. Changed how I debug unfamiliar tools.
Reading source code was expensive. Open the repo, hunt for the right file, trace a call chain, lose an hour. So we read docs instead. Docs were the shortcut.
An AI agent removes the cost. Grepping a cloned repo takes seconds. The agent finds the handler, follows the path, and quotes the line back. Reading code is now cheaper than reading docs.
I tried it last week while testing Floci and Fakecloud as LocalStack alternatives. I wanted to know their limits — which endpoints are full, which are stubs, which quietly no-op.
So I cloned both repos and told the agent: Source is the source of truth. Check it before you claim anything.
Different agent after that. Straight to handler files, back with answers like “this endpoint is a stub” or “header parsed but ignored.” Specific. Verifiable. Often pointing at a line number.
The shift: docs describe intent, code describes what shipped. When reading code was expensive, docs were the right reference. Now that reading code is cheap, code is. Remember to tell the agent.