Skip to content

PHILOSOPHY.md

Latest commit

 

History

History
49 lines (26 loc) · 2.48 KB

PHILOSOPHY.md

File metadata and controls

49 lines (26 loc) · 2.48 KB
Raw

How to get real results from Claude Code

Why some people get an AI that fixes things, and others get one that suggests things.

Claude Code is not a chat box. It can run commands, read and edit your files, look things up on the web and, with brain-in-a-box, remember you. People who get great results do nothing secret: they act as if the list below is true, because it is.

1. Let it act, one door at a time

"It suggested some steps" versus "it fixed it" comes down to access. Let it run commands and edit files in the folder you work in. Connect a service when you need it. Keep approvals on for anything you can't undo: deleting, paying, sending, publishing.

2. Teach it once

Your brain remembers. Say "from now on…", "never…", "next time…": the rule lands in Profile/lessons.md and applies to every future session. A fix that took an hour becomes one sentence next time.

3. Ask "what would it take?", not "can you?"

Missing a skill? It can write one or install one (~/.claude/skills/). A service it can't reach? There is often a connector or an MCP server for it. Nothing exists? It can write a small script.

4. "Done" means proven

The command ran, the page loads, the file exists, the test passes. Ask for the evidence with every "done".

5. Stuck usually means something is missing

Most "I can't" moments are missing plumbing: a tool not installed, an account not connected, a permission not granted, a version too old. Ask "what exactly are you missing?" and fix that first.

6. When something fails, try differently

Diagnose, act, verify, report plainly. The next attempt takes a different route instead of repeating the first one.

7. Automate one chore at a time

Don't design a grand system. Pick your most annoying repetitive task, automate it end to end, live with it for a week, then take the next. A few months later, a good part of your week runs itself.

8. Plan first, then say "go"

For anything big, ask for the plan, read it, then say "go". It can split a big job or a research question across several helpers working in parallel; you supervise.

9. Ask it about itself

Plain words are enough: a screenshot, half a sentence, any language. And ask the meta questions:

  • "What can you do for me?"
  • "Look at my last sessions: what am I still doing by hand?"
  • "What should we automate next?"

In one sentence: treat it like a capable colleague with access, memory and the habit of proving its work, and correct it on the spot instead of giving up quietly.