I use Claude Code to turn recurring Marketing judgments into Skills a team can inspect, rerun, and hand off.
I get a wide view of the market first, then decide where the next hour deserves to go. That keeps my time for the opportunities that genuinely fit and are worth investing in.
I research the role first, until I understand what the team is trying to get done every day and the language they use for it. Before I introduce myself, I want to know what this job really needs — and who I have to prove what to.
I pin down the picture of the work the hiring manager needs to see, then let AI build the tailored resume for each JD, following the judgment logic I've distilled. The HM shouldn't have to translate my experience — they should see right away how I'd judge, how I'd move work forward, and what I'd do well once I'm in.
After AI drafts the tailored resume, I verify the version, the data, and every commitment against real experience — then I submit it myself. AI handles the speed; I stay responsible for what's true: whatever actually ships has to be correct, credible, and genuinely me.
Submitting is only the first touch. After that, I go meet the people worth knowing — it deepens my understanding of the work, and makes me visible to the right people.
When results miss expectations, I first locate where the deviation happened, then figure out where human and AI judgment fell out of alignment. I want the same class of mistake to stop repeating — so every cycle starts from a stronger baseline.
Johanna FanI set the brief, route the right expert, read the funnel, diagnose the constraint, and redirect the next move.
I run this live job search as a closed loop: one brief carries judgment forward; results come back to sharpen the next round's targeting, proof, or workflow.
Explore all public Skills ↗A rule starts as a judgment that lives only in my head, run on feel. Through repeated conversations, AI pushes me to say it clearly — the context, the criteria, the boundaries, the definition of done. Then it rides on real cases: every case keeps its decision and its reasoning; when a judgment drifts, the results correct the logic. Only once the pattern holds does it get wrapped as an agent Skill a team can rerun and hand off.
Judgment written out of my head can be inspected, handed off, and scaled.
State the intent and the why — why this matters, what a miss costs, what good looks like — and the way a model unfolds the work is often more fitting than any steps written for it. So my rules fix what good means first: the goal, the reason, the evidence, the boundaries, the acceptance bar. Only hard sequence dependencies, external release calls, and zero-exception gates get written as fixed steps. The finer the steps, the tighter the system is bound to today's model; workflows that carry intent simply get better when the model upgrades.
Clear intent travels across model upgrades; over-scripted instructions do not.
At runtime, AI produces fast and repeatedly along my decision logic — research, drafts, comparisons, structuring — and it watches more details than one person could hold alone. The outputs return to my hands for the call: I set the brief, handle the exceptions, and keep external release and high-impact decisions mine.
The attention it frees goes back to people — the messages that matter, I write myself.
AI can fail like broken tracking on a dashboard that keeps rendering: silently, fluently, plausibly. So every gate has a catch: before a run, a Skill gets only the complete source context I curated — missing context stops the run; after a draft, a second AI reviews it for accuracy without ever seeing my reasoning; when something slips through, the rule is what changes — the new rule must beat the old one on the same regression cases before it ships; and what no gate can settle surfaces as an exception for a human call.
Context grounded → claim checked → exception surfaced.
Every outcome goes back against the original success criteria and is attributed on one fixed definition. Only patterns with a denominator — comparable, and recurring — change the targeting, the proof, the follow-through, or the workflow. A single result never changes a rule, and one win never becomes the method.
A change isn't learned when it's written — it's learned when the next cycle gets more accurate.
