AICITED

[11] · Experiment · 13 Oct 2026 · 8:25

Claude Code Skills vs CLAUDE.md vs Hooks (I Tested All 3)

8 sources · 8 linked7 chapters
Watch on YouTube ↗Jump to sources

§1 Summary

How to use Claude Code skills, and what we found when we tested where a rule actually reaches. Same rule three ways: the prompt, CLAUDE.md and a skill. The main session obeyed all three. The built-in Explore and Plan agents never saw CLAUDE.md, and only a hook enforced the rule.

§2 Key findings

  1. 01The main session obeyed all three.
  2. 02The built-in Explore and Plan agents never saw CLAUDE.md, and only a hook enforced the rule.
  3. 03Prompt 3/3 · CLAUDE.md 3/3 · Skill 3/3 · Plan mode 3/3 · Explore agent 0/3 · Plan agent 0/3 · general-purpose agent 3/3 · hook only 3/3 (blocked, then passed)
  4. 04Re-run on Claude Code 2.1.296 (10 Oct 2026, 24 runs): the same results.

§3 Our test

  • OUR TEST (Claude Code 2.1.281, 6 Oct 2026, 24 runs, every log kept)
  • Re-run on Claude Code 2.1.296 (10 Oct 2026, 24 runs): the same results. The main session added the code word above a helper's NONE reply in 5 of 6 runs on 2.1.281 and 3 of 6 on 2.1.296.
  • Script and results for both versions: https://gist.github.com/fayaz1010/901f41f5c205bd29e23c3de31d9ed05d

§4 Sources

As published: “SOURCES, checked 6 Oct 2026”

  1. Extend Claude with skills (Claude Code docs)code.claude.com
  2. Create custom subagents (Claude Code docs)code.claude.com
  3. How Claude remembers your project (Claude Code docs)code.claude.com
  4. Automate actions with hooks (Claude Code docs)code.claude.com
  5. Hooks reference, onFailure (Claude Code docs, checked 11 Oct 2026)code.claude.com
  6. Portal by Spotify cut my Claude Code token usage by 90% (Spotify Engineering, 3 Sep 2026)engineering.atspotify.com
  7. Boris Cherny on X, 1 Feb 2026x.com
  8. ep013 lab: one rule, three ways (Claude Code 2.1.281, 6 Oct 2026)gist.github.com

Narration is an AI clone of the creator's own voice. No generated images of people.

Published YouTube description, verbatim (snapshot)
How to use Claude Code skills, and what we found when we tested where a rule actually reaches. Same rule three ways: the prompt, CLAUDE.md and a skill. The main session obeyed all three. The built-in Explore and Plan agents never saw CLAUDE.md, and only a hook enforced the rule.

CHAPTERS
00:00 Cold open
00:45 Same rule, three ways
02:02 How to write a skill
03:34 The helpers never see it
05:44 When a rule has to hold
07:19 How to set it up
08:08 The short version

OUR TEST (Claude Code 2.1.281, 6 Oct 2026, 24 runs, every log kept)
Prompt 3/3 · CLAUDE.md 3/3 · Skill 3/3 · Plan mode 3/3 · Explore agent 0/3 · Plan agent 0/3 · general-purpose agent 3/3 · hook only 3/3 (blocked, then passed)
Re-run on Claude Code 2.1.296 (10 Oct 2026, 24 runs): the same results. The main session added the code word above a helper's NONE reply in 5 of 6 runs on 2.1.281 and 3 of 6 on 2.1.296.
Script and results for both versions: https://gist.github.com/fayaz1010/901f41f5c205bd29e23c3de31d9ed05d

SOURCES, checked 6 Oct 2026
Extend Claude with skills (Claude Code docs): https://code.claude.com/docs/en/skills
Create custom subagents (Claude Code docs): https://code.claude.com/docs/en/sub-agents
How Claude remembers your project (Claude Code docs): https://code.claude.com/docs/en/memory
Automate actions with hooks (Claude Code docs): https://code.claude.com/docs/en/hooks-guide
Hooks reference, onFailure (Claude Code docs, checked 11 Oct 2026): https://code.claude.com/docs/en/hooks
Portal by Spotify cut my Claude Code token usage by 90% (Spotify Engineering, 3 Sep 2026): https://engineering.atspotify.com/2026/09/portal-by-spotify-cut-my-claude-code-token-usage-by-90
Boris Cherny on X, 1 Feb 2026: https://x.com/bcherny/status/2017742745365057733
ep013 lab: one rule, three ways (Claude Code 2.1.281, 6 Oct 2026): https://gist.github.com/fayaz1010/901f41f5c205bd29e23c3de31d9ed05d

Narration is an AI clone of the creator's own voice. No generated images of people.

#ClaudeCode #AIcoding #DevTools