Five Anthropic guides, one broken Claude setup: how a 250-line CLAUDE.md, a 400-line skill and four agents shrink

The five fixes

1. Give the whole job in one message. Per the Opus 5.5 prompting guide as relayed, the model does best given the complete task up front and left to run. The speaker turns four step-by-step messages into one: the goal, who it is for, what good looks like, the file, and when it is done. Spend your words on what a new hire could not guess (your client, numbers and rules). Delete lines like "double-check everything and think carefully": the guide says the model already catches its own mistakes, and when he removed a think instruction in a chat app replies started sooner with no clear quality decline.
2. Cut the context to gotchas. The Claude Code engineer's context article (the source of the "over 80% of the system prompt removed" figure) says too many rules were over-constraining and clashing, such as "leave documentation as appropriate" against "do not add comments". Old rules become judgment (a hard "comments one line max" became "match the code around it"); loading everything up front becomes loading only when needed; examples can constrain exploration (a to-do tool description of about 9,100 characters became one short line); and Claude now saves relevant memories itself, so CLAUDE.md need not be a diary. Keep what Claude would get wrong without being told, such as "the revenue column doesn't include GST". Drop "write in a professional tone". The speaker says /doctor applies these practices to your skills and CLAUDE.md files.
3. Split the skill into layers. At startup Claude loads only each skill's name and description, reads the main file when the skill is relevant, and reads other files only as needed. The main file becomes a contents page that says when to use it and the steps, pointing to a style guide, glossary and example reports. Keep links one level deep, and write a description that says what the skill does and when to use it.
4. Pick effort per job. Start at medium (the Opus 5.5 default) and keep the top levels for work where you have measured a gain. His plan for the report: formatting and email on low, the write-up on medium, the number check on high, because missing weeks and mixed currencies are the hidden edge cases. See our separate page on the effort test.
5. Fold agents into one skill. The October 22 sub-agents guide, as relayed, says a single agent with skills frequently beat both a team of sub-agents and one giant prompt, and one company found the single agent about twice as fast. Merge agents that always run together, call an agent as a tool, or turn it into a skill plus tools. Keep a sub-agent only when you can measure that its context hurts the main agent, when jobs run in parallel with fresh context, or when one agent has too many tools or clashing instructions. In his report only the checker stays, for black-box verification (checking work without seeing how it was made).

The result in numbers

• Four messages became one.
• CLAUDE.md went from about 250 lines to about 35.
• The skill went from about 400 lines that all loaded every run to about 80, plus three files that open only when needed.
• Effort went from max on everything to a level per job; four agents became one main agent plus a fact checker it calls when needed.

Test before you trust it

The sub-agents guide, as relayed, says even 10 golden examples is enough to start (for the report: 10 past weeks where you liked the result). Run each more than once, because in Anthropic's own test one problem appeared in only one run out of two. Anthropic's confidence in its 80% prompt cut came from coding tests that did not get worse.

Checked against Anthropic's skill docs

From Skill authoring best practices, read 2026-10-30.

• "Keep the SKILL.md body under 500 lines for optimal performance" and split content into separate files as you approach it. The video's 400-line skill was under the limit; its problem was loading everything every run.
• At startup only each skill's name and description are loaded; Claude reads SKILL.md when relevant and other files as needed. SKILL.md should work like a table of contents.
• Keep references one level deep from SKILL.md: with nested references Claude "might use commands such as head -100" to preview rather than read the whole file.
• The description should say what the skill does and when to use it, in the third person, and Claude uses it to choose among potentially 100+ skills. "Helps with documents" is the docs' example of a bad one.
• Not checked: the contents of the five guides and the 22 October sub-agents guide, and the /skill doctor command the speaker mentions for skill libraries.