Automation & Agents

Stop Asking AI "What Should I Do": Feed It Your Records, Not Your Story

When you ask AI "should I do X", it almost always agrees with you — because you're feeding it your self-narrative, not your actual behavior. This note flips that: hand it three weeks of work logs, spending records, and your calendar, set one rule, and walk away with a personal operating system you'll use for years.

09/17/2026 ｜ 3 min read

One Core Idea

The quality of AI’s advice depends on whether you feed it a narrative or evidence.

Describe your problem, and it can only extend your narrative — so it always agrees with you. Give it records — what you actually did, where your money and time actually went — and it finally has material to see patterns you can’t see yourself. The whole method rests on one rule:

When what I say and what I do disagree, go with what I did — and treat the disagreement itself as a finding.

How to Do It (one weekend, 2–4 hours)

Step 1 — Inventory your evidence

Gather whatever real records you already have, the rawer the better: work/project logs, git history, budgeting app exports and subscription lists, calendar exports. You can skip this and go straight to the interview if you don’t have these, but the more real the data, the sharper the output. Sensitive details can be anonymized (write “Company A” instead of the real name, magnitude instead of exact amounts) without breaking the analysis.

Step 2 — The opening prompt

I want you to understand who I actually am and produce two documents I can use
for years: a "Personal Decision Constitution" and a "Blind Spot Audit."

Rules:
1. Read the data I provide first, then interview me. Interview in batches,
   numbered, marking which ones I can skip, at most two rounds. Prioritize
   questions about what actually happened (decisions made, regrets, recurring
   situations, where time and money actually went) over what I claim to prefer.
2. When what I say and what I do disagree, go with what I did, and treat the
   disagreement itself as a finding.
3. Tag every judgment with an evidence level: [FACT] (from data I provided),
   [INFERENCE] (derived from behavior), or [GUESS] (insufficient evidence).
   Flag or ask when unsure — never fabricate.
4. You are not my cheerleader. The constitution must be able to stop me when
   I'm being impulsive; an audit that's all praise has failed. Every problem
   you flag needs at least one actionable next step attached.
5. Every rule must be concrete enough to follow literally, with one positive
   and one negative example. Vague language like "stay balanced" doesn't count.
6. Finish one document fully before starting the next.

Step 3 — Answer the interview honestly

It will ask about money, time, regretted decisions, recurring situations. A rough true answer beats a polished, face-saving one.

Step 4 — Check both documents against a skeleton

The Personal Decision Constitution needs at minimum: priority/conflict-resolution rules, fixed conditions for saying yes and no, domain-specific criteria (including a 3–5 question checklist for new opportunities), state clauses (what’s forbidden when tired/late at night/just rejected), and an expiry clause (when the document stops applying and how often to review it).

The Blind Spot Audit needs at minimum: recurring patterns (with evidence), unverified assumptions, single points of failure, over/underestimations, and blind opportunity zones.

Step 5 — Adversarial review (the step most people skip, and the one that matters most)

Open a brand-new conversation and paste the finished documents plus this:

You are an adversarial reviewer, unaffiliated with whoever wrote this. Hunt
for four kinds of defects, ranked by severity:
1. Flattery over help: praise or softened language diluting a judgment that
   should sting
2. Overconfident with insufficient evidence: something tagged [FACT] that's
   actually an inference; a single incident generalized into a "pattern"
3. Unenforceable rules: vague criteria ("in moderation," "reasonable"),
   missing numbers, two rules that contradict each other
4. Internal contradictions: the same threshold stated differently in
   different sections
For each, give one line on how to fix it. Finish with: would this document
actually stop this person when they're being impulsive?

Checking your own output in the same conversation doesn’t work — the AI protects what it just wrote. In my own run, this pass caught 12 issues, two of which would have made the entire rule set unenforceable.

Step 6 — Ask for an early-warning table

Format: signal → what it means → action. The hard rule is signals must be observable (account balance, file timestamps, how many slots behind schedule) — never “feeling tired” or “feeling anxious.” This way, any AI in any new conversation can run the check for you without relying on your own self-awareness in the moment.

Step 7 — Store it and actually use it

Save it in whatever note system you already use. Before any real decision, open a new conversation, paste the constitution, and require “answer by citing clause numbers” — an answer that doesn’t cite a clause counts as not having read it. Schedule a review every six months: hand the AI the constitution plus the last six months of records and ask for an adversarial re-audit — outdated clauses and evidence of you routing around your own rules.

Pitfalls You Will Hit

1. Asking the AI to self-check in the same conversation — it will protect its own output. Review must happen in a fresh conversation.
2. Accepting rules with no numbers — “don’t make big decisions when tired” is unenforceable; “slept under 5.5 hours → routine tasks only” is.
3. Forgetting the expiry clause — after a job change or income shift, stale rules are more dangerous than no rules.
4. Treating the audit like fortune-telling — a [GUESS]-tagged section means the evidence is thin; the right move is to go get more data, not to accept or reject it wholesale.

One-Line Summary

The most valuable thing AI can do isn’t answer your questions — it’s show you the data you’ve been avoiding, as long as you feed it evidence and send in a second copy of itself to catch the flattery.