I asked Claude Code for the same 3D engine twice with the teardown skill.
From my sentence alone: a generic engine.
With a museum photo: Claude read the carburettor label, found it was a Moto Guzzi Stornello 125, looked up its real size and copied it part by part.
Start with 1 procedure you paste every week: .claude/skills//SKILL.md, description line first.
Only install skills from sources you trust.
Docs: https://t.co/O8ERtSCj8E
h/t 0xsojalsec on Instagram
I put the same 25 procedures into CLAUDE.md, then into 25 skills.
CLAUDE.md added 34,388 tokens to every message.
The skills added 1,184.
That's 29 times less, for the same steps.
So fix it in your request. Say release notes, or say changelog.
And write each description in third person: what it does, and when to use it.
"Writes release notes for users. Use when asked for release notes or what is new."
Your CLAUDE.md gets added to every message.
I moved the same 25 procedures into 25 skills: about 34,000 extra tokens per message became about 1,000. Claude only reads each skill's one-line description until it needs the rest.
Same page, same selector, scraped twice with Claude Code. One run sent Claude 29,758 characters, the other 6,742. The difference was 2 words in the prompt.
Claude picked the right Scrapling tool every time. The waste was the format: it asked for HTML. Add "as markdown" next to the selector: same books, under a quarter of the text.
The limit: "as markdown" on its own didn't do it. Claude still asked for HTML 3 of 3 times. With the selector in the same prompt it was markdown 3 of 3.
Free, BSD-3-Clause: https://t.co/nWEtMxNTjS
h/t @bestapps.ai
I scraped the same page twice with Claude Code, same selector both times.
One run sent Claude 29,758 characters. The other sent 6,742.
The difference was 2 words in the prompt.
Setup in Claude Code:
pip install "scrapling[ai]"
scrapling install
claude mcp add ScraplingServer "$(which scrapling-mcp)"
Restart Claude Code and ask for the data, as markdown.