待翻譯：Captaining your AI: new human-agent UX paradigm

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

Michael Carroll Aug 09, 2026 This is part 2 of 2 on the Captain Strategy. Part 1 argued that we’re all using AI to become the bosses we hate — micromanaging assistants instead of building functions with defined inputs and reviewable outputs. Now, how to get to AI functions? It starts with changing the entire way you interface with agents. First, your bi-weekly dose of AI news (as if you aren’t inundated with it already): Here’s a new news category I’m sure you will see more of: Add AI to Old, Boring Thing & Turn it Into News The boring thing: A hedge fund overleverages itself and implodes. Why it’s boring: Yes, these are a lot of fun to watch, especially when the hedgie in question is young with a very chiseled, instantly unlikeable face. But we basically get a new story of “hedge fund takes on too much debt and firesells everything“ every 3-5 years since the 90s. It’s more predictable than presidents campaigning on tax cuts. How AI got added to it: The guy making the bets wrote essays on AI (like 90% of the Internet today). He bought heavily into the AI bubble (like 90% of Wall Street today). He used to work for an AI company (which, if you believe the press releases, are 90% of all companies). Apparently all of that equals “big AI scoop”. Verdict: 🙄 but with a pleasing dose of schadenfreude HBR: AI agents are cheaper and faster than assistants I wrote about this last week. Maybe you didn’t believe me? Well Harvard Business Review & Perplexity folks helpfully crunched the data and put it in a chart for you. Worth reading the whole article! NYT primer on tokenomics You know a tech trend is going mainstream when the NYT writes about it. Somehow, though, every “expert” quoted in the article seems focused on convincing how hard of a problem determining value of AI is. You can practically hearing them sharpening their knives as they position themselves to charge enterprise exorbitant “AI tokenomics” consulting fees. Well, not to be left out, here’s my shameless plug: Coolhand Labs gives you basic token/cost calculations you can stand on for free. No need to spend six figures on consultants. We will give it to you for free. (You are welcome.) Now, let’s talk about how to manage your AI workloads like a captain! What a captain actually does A ship’s captain doesn’t do every job on the ship. In fact, you could be forgiven for wondering if they do any job at all. Captains DON’T: ❌ stand in the engine room adjusting valves ❌ personally trim the sail ❌ take the wheel and steer... unless, of course, the ship is sinking. Better put: if you are doing every single job on a boat to make it sail, you aren’t captaining a ship: you are rowing a canoe. In my last post, I talked about why people aren’t finding productivity or satisfaction with agents: they’ve used AI to go from rowing the canoe to shouting realtime commands at whoever’s holding the paddle. It’s a management anti-pattern that’s been around so long that Dr. Seuss parodied it sixty years ago, in I Had Trouble in Getting to Solla Sollew: But “Go left, go right, all day and all night” is exactly where we end up when we use chatbot interfaces! Nobody in that picture is a captain — and the chap furnishing the brains isn’t getting to his destination any faster for it. Spoiler Alert: None of the people in that book ever get to Solla Sollew. An actual captain’s job is to provide direction to the crew. Sometimes they give it in realtime — navigating a storm — but typically they give it by reviewing the crew’s reports at the end of the day. The Captain Strategy asks you to approach using AI the same way: instead of parking yourself on top of every turn the AI takes, move to reviewing the AI’s final outputs — or, at most, the one genuinely critical gate the AI can’t clear on its own, because it lacks information only a human has or because something must be verified before the rest of the work can safely proceed. Captains SHOULD: ✅ set the course ✅ review the outcomes of the crew’s daily work ✅ give feedback and, if necessary, dig into the details of systemic problems ✅ with any spare time, hunt down inefficiencies across the whole operation (always easy to find if you bother to look) This isn’t a workflow tweak. It’s a re-architecture of the job’s entire UX — from “micromanaging a worker” to “captaining a process.” What’s holding most people back from this? It’s often not that they love micromanaging too much — it’s that they don’t have their AI processes set up in a way that can enable this. So let’s talk about how you get there. How to build it Here’s a rough rundown of how to build your AI workflow using the Captain Strategy: 1. Never “prompt” an AI Turn your AI process into a function — don’t build it as an assistant. In simpler terms? NO CHATBOTS. Instead, build AI like a function. It should: be triggered by an event — an inbound email or call, a time-based activation, whatever wake up with a “baseline context” — i.e., your prompt get the data it needs from available sources — API & tool calls perform some actions — clean the records, draft a report then, finally, queue the results up for human review 2. Start from the quality gates you already have A mature process already has places where work gets sampled, escalated, or sent up for review. THAT is where review should happen. What’s important to remember is that every quality gate is different, and they evolve with the business and your confidence. An immature process may need a gate where every AI output lands in a “draft” state and gets human review before it’s finalized or the rest of the process moves forward. But as the AI function improves and confidence increases, the gates should evolve too. A sign of a function that’s getting to maturity: a human doesn’t need to look at every output — just a 20% sample. At the “sampling point”, outputs only need review after the fact, because most of the review is tweaks. For some processes, confidence will get so high that the quality gates disappear altogether, or remain only as a way to keep the AI function in sync with the evolving processes around it (”we don’t need X statistic in this report anymore — exclude it going forward”). 3. Create a “single pane of glass” for reviewing AI outputs... and time-box that process Rather than sticking with the old model of disparate dashboards, create a single interface for reviewing all the key outputs in a time-boxed way. Think of it like the captain meeting with their first mate and lieutenants in the morning and at the end of the day: a single period set aside to review the important items, give directions, and send everybody off to make it happen. The time-boxing is the part everybody skips, so make it explicit — cap reviews at two hours a day. I call this the Captain Shift: the shift (in both senses) where you’re on deck reviewing the crew’s work, with a hard stop. Everything outside the Captain Shift is reserved for the focus work that moves the whole ship forward — digging deep into a persistent quality issue, or realigning with the latest directions that came in by carrier pigeon. If reviews swallow the whole day, you are still micromanaging; you’ve just rebuilt the button-pushing job with a nicer dashboard. 4. Don’t just pass/fail. Edit and explain What distinguishes a great manager from a lackluster one? The great manager doesn’t just make their team work — they help their team understand and improve. If a captain just passes good results and fails bad ones, chances are their team will begin to understand what’s good, what’s bad, and what to do better next time. Sometimes. Maybe. But a captain who explains the *why* of their decisions gives the crew a framework — not just for what’s good in this situation, but for the standards of good and the goals of the task. AI functions level up the same way: “our goal is to preserve the anonymity of the user” will always beat “don’t put in birthdates.” Closing the loop: captain reviews are possibly the most valuable data your company has Those markups — the edits, the “do it this way, and here’s why” — are the highest-signal feedback your AI process will ever receive. It’s a domain expert — YOUR COMPANY’S domain expert — telling you, in specifics, exactly where the output falls short and why. And that expertise? It’s what gives your company its competitive edge. Most companies don’t win by doing something nobody else can do. They win by improving how they do it faster & better than everybody else can. Captain reviews reinforce your existing moat. But to put water and sharks in that moat, you need to feed the reviews back into the AI function. Captain edits need to become prompt fixes. They need to kick off tool investigations — *why does the report pull the wrong number in this one section?* They need to launch product and efficiency discussions upstream of the AI entirely. These are the things that make an AI process reproducibly better, not better-this-one-time. Some teams start by running that translation by hand — the captain marks up the output, an engineer converts the markup into prompt and tooling changes. But that quickly eats the very hours the Captain Shift freed up. Automating it is the bulk of what we do at Coolhand Labs: the dredge work (the pun stays) of turning captain feedback directly into prompt, tool, and process improvements. NOTE: We also open-source all of Coolhand’s interfaces and tools for building self-improving AI feedback loops, in case you’d rather just build this yourself (no judgement). Either way, your goal is the same: a crew that is constantly improving its work patterns & outputs in response to the captain’s reviews. A process that absorbs feedback compounds. One that doesn’t just repeats its mistakes at fiber-optic speeds. One final note: reviews can be fun! One tactical note that matters more than it sounds: don’t dump outputs into a big list and ask a captain to go through each one. A list is a chore, and it naturally leads people to just graze past looking for big failures, rather than spending time on a single item and asking how it could be the best it can be. The better way: Serialize the outputs. One output at a time, in sequence, no skipping — and time the queue so it fits inside the two-hour Captain Shift. Then get smart with sampling, because sampling is how you scale the captain’s attention: review everything early, dial mature processes down to a small fraction, and oversample whatever’s new — a fresh prompt, a new tool — so you can actually see how the change performs. You can even bake in an A/B test, if you like: half the sample from the new pattern, half from the old, head-to-head. Like gamifying things? Throw the captain a trophy for particularly insightful reviews or review streaks! Also important to notice? The purpose of the review is different, depending on the maturity of the AI process: ➡️ On new processes it’s quality gating — catching real errors before they ship. ➡️ On mature ones it’s alignment — catching the moment the process drifts out from under a business that’s moved on. It helps to label each output with its workload’s maturity — automatable, if you have quality scoring in place — since reviewers give better feedback when they know whether they’re gating quality or checking alignment. Captain Shifts are MUCH bigger than “better ops AI” You could file all of this as the staffing chapter of claudeshoring. It’s true: I don’t know how you’d run a claudeshored process well without it. But I don’t want to undersell how valuable captain shifting is: The Captain Strategy is a general pattern for how humans and AI should work together Time and again I see people claiming the AI isn’t good enough to deliver ROI. But the issue isn’t the AI — or the people using it — it’s the bad practices their human-to-AI interfaces lead them into. Captain shifting solves that for ma [truncated for AI cost control]