Labor into skills. The work you used to do by hand has a system that does it.

~/.claude/skills/
└── marketing-helper/
    ├── SKILL.md          ← ~12,000 tokens
    │                       "anything marketing related"
    ├── ig-content.md
    ├── email-templates.md
    ├── ad-account-stuff.md
    ├── retention-email-old.md
    ├── retention-email-new.md
    ├── content-calendar.md
    ├── caption-ideas.md
    └── misc/
        └── ... (47 files)

One enormous "marketing-helper" skill that promises to handle anything marketing-adjacent. Claude reads the full SKILL.md every session that smells like marketing, even when you only asked about pricing. The internal files contradict each other in places. Claude doesn't know exactly when to invoke it because the description says "anything related to marketing."

Half the files in misc/ were one-off experiments that should never have made it into a skill. Token cost per invocation: heavy. Confusion cost: high. Sarah opens Claude on a Tuesday and gets vague drafts because Claude pulled in irrelevant marketing context she didn't ask for.

~/.claude/skills/
├── pull-weekly-revenue/
│   └── SKILL.md          ← ~450 tokens, one job
│
├── draft-retention-email/
│   └── SKILL.md          ← ~520 tokens, one job
│
├── audit-ad-account/
│   └── SKILL.md          ← ~680 tokens, one job
│
└── team-brief-from-slack/
    └── SKILL.md          ← ~390 tokens, one job

Four small focused skills. Each does ONE job. Each description tells Claude exactly when to invoke it, so most sessions invoke zero of them and pay nothing. No overlap, no confusion. Sarah uses each one multiple times a week.

When she discovers a new repeating job (say, weekly Google review reply drafts), she adds a fifth small skill instead of stuffing it into one that already exists. The folder grows wider, never deeper. Each skill stays small enough that Sarah can read it in thirty seconds and know what Claude will do when it fires.