Judgment at Scale
The hard part was never doing more work. It's doing more work without diluting the judgment that made the work worth paying for.
Judgment at scale means your best calls, patterns, and standards show up across far more output without turning into generic mush. AI is how you broadcast that judgment, not a replacement for it. Used well, Claude makes it feel like you've cloned your own thinking, not hired a bright stranger with no context. The trap is treating it as a faster junior who writes in nobody's voice. The prize is using it as an amplifier of the specific, opinionated judgment you spent 25 years building.
I missed this for the better part of a year. I used AI to sling out more memos, more first passes, more options. The work was technically correct and spiritually wrong. It lacked the edges I'm paid for. A client finally told me a deliverable "didn't sound like you, I could've gotten this anywhere." That stung, and they were right. I'd scaled typing. I hadn't scaled discernment.
What does "judgment" actually mean here?
Judgment is everything you do that doesn't show up in the raw facts: what you ignore, what you push on, how you trade off risks, which failure modes you worry about, and which standard answers you quietly bin. It's the difference between an experienced partner and a search result. That's the part worth scaling.
There's an uncomfortable part to this for anyone over 40. AI has already flattened access to information. A 28-year-old analyst and a 55-year-old partner can pull the same statutes, filings, and benchmarks in seconds. The gap is no longer who can find the thing. It's who knows which thing matters and what to do next. If you use Claude in a way that erases that distinction, you're doing unpaid training for your own replacement.
How do you encode judgment instead of just asking for drafts?
You stop treating prompts like magic spells and start treating them like operating manuals. The simplest version is a Standards Document: two to three pages you feed into a Claude Project so every session starts with your working principles, not a blank slate. It answers four questions: how you think, what you always check, what you refuse to do, and what good looks like in your world. You're not teaching the model to sound smart. You're teaching it the moves you make when nobody's watching.
- Capture your refusals. Write down the shortcuts you won't take. "Never recommend a vendor I wouldn't use with my own money." "Never give a one-number forecast without showing the range." Those are sharp edges, and a generic model will never infer them.
- Give it three exemplars. Paste in three pieces of work you'd be proud to show a tough client today. Claude learns more from structure, pacing, and emphasis than from adjectives like "strategic."
- Name the lens. Say how you show up. "I'm a sell-side M&A advisor writing for skeptical buyers." "I'm a litigator reading this contract to break it, not to close it." A label like that changes the output more than any style instruction.
- Keep it living. Every time you catch yourself thinking "I'd never say it that way," add one line to the document. Within 20 to 30 iterations, the misses start repeating less.
How do you actually set this up in Claude?
If that sounds like extra work, this is the operational version I use and recommend to clients. You can build it in an afternoon and refine it over a month.
- Create a dedicated Project. One Project per recurring deliverable: Board Memos, Deal Screening Notes, Client Advisory Letters. Don't pile everything into one thread.
- Add your Standards Document. Upload it as a short PDF or pin it as text. Version it ("Screening Standards v1.2") so you remember it's meant to change.
- Attach two or three finished examples. Redacted work you'd actually sign today, labeled so future-you remembers why each made the cut.
- Use a consistent setup prompt. Start each thread with something like: "Work under the attached Standards. Draft a first pass I can mark up, stay inside those rules, and when you're unsure, ask before inventing." Save it so you're not rewriting it every time.
- Close the loop. After you edit, paste your changes back in: "Here's what I changed and why; treat it as an update to the Standards." That's how you compound judgment instead of re-teaching it every session.
On my own work, this took output from "competent but anonymous" to roughly "95% there" over about 20 documents. Crude measure: editing a memo went from around 45 minutes to 10 or 15, because the model already made the obvious cuts I would have made anyway.
What is the "Could I sign it?" test?
The cleanest safeguard I've found is brutally simple. Before anything leaves your desk, ask: if I had to sign this right now, zero edits, would I? Not "is it close," not "could a smart reader infer what I meant." Yes or no. When the answer is no, the interesting question isn't "fix the draft." It's "what exactly makes this un-signable?" Maybe the risk section buries the worst case. Maybe it sounds more confident than the evidence justifies. Maybe it's polished and says nothing new. Those gaps are your judgment, and they belong back in the Standards Document so the next draft closes them.
This does two things. It keeps your name clean, the danger in scaling with AI isn't typos, it's quietly lowering the bar on what your signature implies. And it forces you to codify the felt sense you built over 20 or 30 years into rules something else can actually follow.
A composite from practice
Take a pattern I see often, call him a 58-year-old M&A advisor who used to write every deal memo by hand, roughly six hours each, which capped how many live deals he could run at once. His expertise was real and his throughput was his ceiling. He built a Standards Document capturing how he assesses a target: the three questions he always asks about customer concentration, the red flags he's learned to smell, the way he frames downside before upside. Now Claude produces the structural first draft in his framework, and he spends his hours on the judgment calls instead of the typing. He didn't get faster at writing memos. He got faster at the parts a machine can do, so he could spend more time on the parts only he can. His live-deal capacity roughly doubled, and the quality went up, because he was finally spending his hours where his edge actually lived.
Manual prompting versus encoded judgment
| Dimension | One-off prompting | Encoded judgment (Standards Document + Project) |
|---|---|---|
| What the model knows about you | Nothing; starts cold each time | Your standards, refusals, voice, frameworks |
| Output voice | Generic competent professional | Recognizably yours |
| Your role | Editing a stranger's draft | Directing a draft already in your frame |
| What scales | Volume of output | Reach of your judgment |
| Risk over time | Drift toward the generic mean | Compounds as you refine the document |
| The 40+ edge | Erased; everyone prompts the same | Amplified; your experience is the input |
Where this leaves you
The professionals who win the next decade won't be the ones who produce the most with AI. They'll be the ones whose judgment is encoded so cleanly that the tool becomes a multiplier of their specific point of view. Your experience is the rarest input in the system right now. Don't spend it supervising generic drafts.
Start this week with one deliverable you produce on repeat. Write the two-page Standards Document, your refusals, three exemplars, your lens. Run your next draft through Claude inside a Project, then apply the "Could I sign it?" test. Whatever fails goes back into the document. Do that for a month and you'll have built something no competitor can copy: a tool that thinks the way you've earned the right to think.