5 Claude Code Patterns That Separate Power Users From Everyone Else
Skip if you’re new — these are for people past CLAUDE.md and settings.json. The patterns below are what I see in Claude Code workflows that actually ship, scale, and survive past the first month.
1. Skill-driven workflow forks instead of one mega-context
Most users stuff their CLAUDE.md with 800 lines hoping Claude reads it all. It doesn’t work. The model skims the front, half-attends to the middle, and the back gets context-dropped under load.
The fix: split the monolithic CLAUDE.md into skill files that load conditionally.
.claude/skills/
refactor.md ← loads only when "refactor X" prompts
deploy.md ← loads only when "deploy" prompts
testing.md ← loads only when "test" prompts
db-schema.md
security.mdSay “refactor X” → only refactor.md loads. Say “deploy” → only deploy.md loads. 60–80% context savings, and Claude actually pays attention to the rules in scope.
2. PreToolUse hooks for destructive command gates
Approval prompts don’t scale to background sessions where there’s no human to approve. They block autonomous loops cold.
The fix: a 30-line PreToolUse Bash hook + allowlist file kills every rm -rf, DROP TABLE, git push --force, and chmod 777 -R before the model can fire them.
# .claude/hooks/pre-tool-use/bash-vet.sh
case "$1" in
*"rm -rf "*) exit 1 ;;
*"DROP TABLE"*) exit 1 ;;
*"git push --force"*) exit 1 ;;
*"chmod 777 -R"*) exit 1 ;;
esacHooks run silently, scale to autonomous loops, and you sleep without wondering what your agent shipped at 3am.
3. Subagent task delegation
Stop running one session that does everything. Long sessions hit “compaction drift” — the model loses early decisions as the middle gets compressed under load.
The fix: spawn focused subagents per task type.
- Research subagent → reads docs, explores code, summarizes
- Planning subagent → takes research output, drafts the change plan
- Execution subagent → implements one phase at a time
- Verification subagent → runs tests, checks output, reports back
Each subagent has clean context. Cross-context isolation prevents the drift. Surprisingly cheap given the token cost of one bloated session vs four focused ones.
4. Headless mode + scheduled cron firings
This is the unlock most users haven’t tried.
Anthropic’s new scheduled tasks + a state file in the repo + headless prompts = genuine 24/7 work without an active terminal.
# .claude/schedule.yaml
- name: dependency-bump-friday
cron: "0 14 * * 5"
prompt: "Audit package.json. Open PRs for non-breaking updates."
- name: daily-changelog-refresh
cron: "0 6 * * *"
prompt: "Read git log of last 24h. Append to CHANGELOG.md."Code reviews, dependency bumps, content generation, daily refreshes — all autonomous. The state file holds context across firings so the agent picks up where it left off.
5. Per-repo settings.json with command-level allowlists
Don’t share one global allowlist across projects. Each repo gets its own .claude/settings.json defining safe commands for that codebase.
{
"permissions": {
"allow": [
"bash:npm test",
"bash:npm run build",
"bash:git status",
"bash:git diff *",
"bash:gh issue *",
"edit:src/*"
]
}
}Cold-start onboarding on a new repo drops from 30 minutes of approval-fatigue (“Claude wants to run npm test? Yes. Claude wants to run git status? Yes…”) to zero. The settings.json defines the safety boundary once; every session in that repo respects it automatically.
Where this leads
These five patterns aren’t visible in the official docs. They emerge from running Claude Code at production scale. If you’re past the CLAUDE.md basics and want the next layer — start with skill-driven forks. The context savings alone change how the tool feels.
Bookmark for your next workflow refresh.
Originally posted on X as a long-form article: 5 Claude Code Patterns That Separate Power Users From Everyone Else.