Operator Guide · Chapter 11

Anti-patterns I learned by stepping in them

Eighteen months of running this for myself surfaced about thirty anti-patterns. The fifteen that hurt most are here. If you avoid these, you'll skip a year of pain.

01Treating BRAIN.md like a journal

The temptation is to write paragraphs in BRAIN.md describing what you've been thinking about. Don't.

BRAIN.md is current state, not narrative. If a section is >30 lines, you're storing history that belongs in CHANGELOG.md or MEMORY.md.

Symptom BRAIN.md grows to 500 lines and becomes unusable.

Fix Every quarter, archive anything in BRAIN.md that's no longer current. Resolve completed priorities into CHANGELOG entries. Move durable observations into MEMORY.md.

02Daily logs as journals

Same problem one layer down. Daily logs that read like diary entries — "today I felt anxious about the meeting, then I got coffee, then I started thinking about..." — produce no re-readable signal.

Symptom Daily logs run 1500+ words. Reading yesterday's log to remember where you left off takes 10 minutes.

Fix Enforce the seven-section structure. If a section would benefit from a paragraph instead of bullets, the section is wrong for that content. Move it elsewhere.

03Premature abstraction

You see two vendor invoices that look similar. You build a shared template. Now every vendor invoice has to fit the template, including the third vendor whose invoices have a completely different structure.

Symptom Complex generic systems for things you do 3 times a year.

Fix Don't abstract until you have THREE concrete instances. Until then, write the work out longhand for each instance. The "duplication" is cheap; the wrong abstraction is expensive.

04Migrating everything in a weekend

You buy the kit. Inspired by the structure, you spend a Saturday migrating every document you've created in the last five years into the new vault. By Sunday night, you're exhausted and the system feels like a chore.

Symptom High enthusiasm on day 1, abandonment by day 14.

Fix Migrate as you go. Start writing today's documents into the vault from day 1. Migrate older documents only when you actually need to reference them. Most archives sit unused; migrating them is wasted effort.

05Tracking metrics you can't update

You build a Numbers to Know table in BRAIN.md with 15 metrics. Three weeks later, half of them are stale because updating them takes longer than you have.

Symptom BRAIN.md shows numbers from 6 weeks ago, undermining trust in everything else in the file.

Fix A metric only goes in BRAIN.md if you can update it in under 60 seconds. Anything that requires a query, a calculation, or a third-party login lives in _loop/measure/, with the BRAIN.md value pulled in only when fresh.

Three metrics you update faithfully > fifteen metrics you update sometimes.

06Signal noise

You build 20 loop signals because the chapter on signals was inspiring. Within a month, 8 of them fire constantly because thresholds were too aggressive. You learn to ignore them. Now they don't even surface real problems.

Symptom Morning briefs cluttered with INFO-severity signals you've stopped reading.

Fix Start with 3 signals. Run for 30 days. Tune. Add 3 more. Repeat. Volume is the enemy of signal.

Also: any signal that's been active for 30+ days has become wallpaper. Either solve the underlying issue or kill the signal.

07Letting cron fail silently

You set up a cron job. It works for two weeks. Then macOS auto-updates, Full Disk Access gets revoked, and the cron stops running. You don't notice for six weeks because the daily output it produced was just always there in the background.

I lived this one. My SHIFT4 ingest cron reported "success" for seventeen straight days while writing zero rows to the database. The anon key didn't have INSERT permission on the table; the API returned HTTP 201 (success) but the row never landed. I discovered it only when the downstream dashboard went flat for two weeks. By then I'd lost half a month of operational data.

Symptom Discovering weeks later that a critical cron stopped running. Data gap. Trust degraded.

Fix Every cron writes a watchdog event at end of run. A cron-doctor job (every 4-6 hours) checks that every monitored cron has reported recently. If any haven't, the doctor pages you.

This single discipline saves more pain than almost anything else in the harness.

08Customizing on day 1

You see the default category names and immediately want to change them. "Resources" feels weird; you want it to be "Vendors." "Relationships" feels too broad; you want to split it.

Symptom Vault structure constantly being reorganized. Files moved repeatedly. AI tools can't navigate consistently.

Fix Run the defaults for 30 days before customizing anything. Most customizations people propose on day 1 get reversed by day 21 because the defaults turn out to be reasonable.

The customizations that DO stick are usually small (one category renamed, one subdirectory added). Major restructures don't survive contact with real work.

09Skipping the five-minute session-end discipline

The discipline is BRAIN.md update + CHANGELOG.md entry at session end. Five minutes.

You skip it once because you're tired. Then twice. Then BRAIN.md is two weeks stale, you don't trust it, you start consulting it less, and the whole brain layer atrophies.

Symptom BRAIN.md hasn't been updated in 10 days. Daily logs exist but don't roll up into anything. The harness becomes a write-only system.

Fix Treat the five-minute session-end discipline as non-negotiable. If you can't spare five minutes, you didn't actually finish the session. Schedule it into your end-of-work routine the way you'd schedule a shutdown checklist for any other operational system.

The harness depends on this. Skip it and everything else fails downstream.

10Generating content with AI and accepting the first draft

You ask the AI to write a vendor email, generate a brief, draft an SOP. The AI produces a draft. It's fine. You send it.

Three months later, you notice that your "voice" has drifted toward generic corporate-AI-speak across everything you publish. You can't quite tell why your writing feels less like you.

Symptom AI-generated content shipping without rejection-pass.

Fix VOICE.md is the law. Every AI-generated draft gets read against VOICE.md before shipping. If it doesn't match, reject the draft and re-prompt with stricter constraints.

This is especially true for: emails to customers, social media posts, anything that builds your public brand. The cost of drift is invisible per-instance but compounds across hundreds of touches per year.

11Treating the vault as backup

The vault holds your operating reality. It is not your backup.

If your drive dies and the vault is the only copy, you lose everything. Cloud sync (iCloud, Dropbox) is not a backup — sync errors propagate.

Fix Set up two independent backup channels. Time Machine + Backblaze. Or rsync to a second drive + cloud snapshot. Verify backups every quarter by restoring a random file to a test location.

12Building automations before you have the manual habit

You see how powerful cron + Ollama could be and try to automate everything. The morning brief, the inbox classifier, the loop signals, the cron-doctor — all on day 2.

You haven't yet built the habit of reading the morning brief, acting on it, updating BRAIN.md based on it. The automations produce output but you don't consume it.

Symptom Lots of generated files; same chaotic operating reality as before.

Fix Build the manual habit first. Write the morning brief by hand for two weeks. Once you reliably read and act on it, automate. Same for any other automation — manual habit precedes automation.

The automation amplifies a real habit. It cannot create one.

13Confusing the vault with productivity

The harness is not a productivity system. It's an operational system.

Productivity systems try to make you do more things faster. The harness tries to make sure the things you do compound — that decisions you make stick, that observations turn into patterns, that one-off heroics become repeatable processes.

If you're optimizing for "how many tasks can I close per day," the harness will disappoint. If you're optimizing for "what does my operation look like in 12 months if I keep running it," the harness compounds.

The right mental model is closer to "the operating manual for a small private equity firm" than "Getting Things Done."

14Sharing the vault prematurely

A partner, contractor, or family member asks to see the vault. You share read access.

They see a system that took you 12 months to build, presented as if it sprang into being fully-formed. They feel intimidated. Their participation degrades.

Symptom Shared vault becomes hostile turf.

Fix Build handoff packages for collaborators, not full vault access. A handoff package is a thinned, simplified version with only what the collaborator needs. (See chapter 12 on the 30-day curve, where this is covered for the "multi-domain + handoff" tier.)

When I was ready to share part of the vault with my wife, I didn't share the vault. I built her a thinned version — six files instead of six hundred, Spanish-first, with a "what NOT to touch" doc and a single starter command. The whole package was designed against the failure mode in this anti-pattern. She runs it. The full vault would have intimidated her into never opening it.

Your partner doesn't need to see your tax filings. Your contractor doesn't need to see your investment thesis. Filter.

15Treating the harness as a project

The harness is not a project. Projects end. The harness is the operating layer.

If you treat it as a project — "I'm setting up my harness this month" — you'll declare it done at some point, stop maintaining it, and watch it decay.

Treat it as discipline. Like brushing your teeth: never done, never abandoned, runs at the same cadence for years.

The meta-pattern

All fifteen anti-patterns share a root: trying to do too much, too fast, without earning the right discipline.

The harness rewards patience. The first 30 days install habits. The next 60 days compound them. The next year produces real leverage. There's no shortcut. People who try to skip steps abandon the system; people who follow the curve get the compounding effect.

Next chapter: the 30-day curve, and how to survive it.