Three Months With Claude Code: What 35 Jobs and 26 Killed Ones Taught Me Bottom line up front. AI support cut 90% of my writing time. Three months in, I built 35 automation jobs and killed 26. Volume alone doesn't make money — that part I confirmed the hard way. 3 months in 35 jobs active 26 killed 90% writing time saved Start with the tangible stuff. This is the automation surface I've built up over three months. Every morning, launchd runs the lot without me touching anything. 35 LaunchAgent jobs (26 killed) 12 project folders (5 blogs + α) 3 custom skills (only SKILL.md loads) 19 slash commands (/run-daily & co.) On top of that: 13 CLAUDE.md files (1 main + 12 per-project) and 3 hook scripts. Every job ships with a wrapper, plist, and log file — so the actual file count is well past 100. In Claude Code, context tokens are the bill. After three months, these are the five habits that stuck. /compact — the moment the context warning fires. When the alert pops, hit it immediately. Keeps the gist, drops the rest. Split out agents — separate roles for detail and deep work. Main session is the coordinator; details go to a sub-agent. Keeps the context from getting polluted. /clear — every time the task switches. Going from blog work to infra debugging? Clear, no exceptions. Keep CLAUDE.md slim — only the must-reads. It auto-loads every session. Push the rest into separate files and reference them only when the task calls for it. Skill split — load only SKILL.md. My three custom skills sit in their own files. The full text stays out of context until needed. Out of 19 commands, 3 skills, and a pile of plugins, these are the three I genuinely reach for daily. Rank Tool How ROI A · Daily publish /run-daily ×5 per-blog .claude/commands/run-daily.md edit one line → all 5 blogs pick it up B · Writing quality blog-style-guide 9 base patterns + 4 evolved patterns in one file tweak the guide once → every post follows the new rule C · Design frontend-design installed 4/26, generates v2 visual components 3–5 polished components per post, on demand The frontend-design plugin in particular changed what I can ship. Before it, I was hand-rolling CSS and the readability suffered. Now Hero, VS, and impact blocks come out clean in one pass. Eleven days in, the lift in visual quality is immediate. "I assumed automated posts would mean automated income. Three months later, it's the opposite." Root cause: automation cuts repetition cost. Revenue is built on trust. You can ship 5 posts a day, but if every one reads like a textbook, readers click and bounce. Look at the asset count again. 35 active + 26 killed. Those 26 are the ones I built on the assumption that volume alone would pay off. Early assumption (26 killed jobs): "Auto-publish daily" Generic / textbook tone Conversion flatlined After the pivot (35 active jobs): Experiment logs + lived experience Drafts → my own review Reads trending up Automation cuts repetition cost. Revenue compounds from trust. Two completely different mechanisms — I had to kill 26 jobs to internalize that. Volume without trust is worse than zero posts. Two different contexts auto-load every session. Mix them up and things break. Criteria CLAUDE.md (project file) Memory (auto-memory) Holds Operating rules, publish rules, incident logs Meta info, persona, where external systems live Update freq Often (every time something breaks) Occasionally Trust High — source of truth for ops Supporting — meta only The split rule is simple. "This needs to work in the next session" goes in CLAUDE.md. "Who the user is, where they work" goes in memory. Don't trust memory for operating rules — always pin them in CLAUDE.md. I caused several incidents over three months by getting this wrong. One more thing: don't put it in either place. Concrete values (cash balances, prices) or transient state. They go stale fast — push them to a file or a database instead. "You still need a human's lived experience. Trust beats volume. But AI support did cut 90% of my writing time." — Eddie, 3 months with Claude Code This post is the index. Each topic gets its own case-study post next. Token saving deep dive — each of the 5 habits, in practice Slash commands deep dive — how /run-daily differs across 5 blogs 3 custom skills deep dive — why I built them, how they're shaped Running CLAUDE.md — how I let incident logs accumulate frontend-design plugin — the 30-day review Hooks, memory, agents — the rest of the stack Productivity and trust aren't substitutes. AI buys back the time. The human fills in the trust. Disclaimer: personal 3-month retrospective. No ads, no affiliates. Originally published at jessinvestment.com.