I’ve been preparing a talk on getting real, reliable work out of AI, and I went down a deep research rabbit hole along the way. I read peer-reviewed studies, engineering write-ups from companies building these tools, and a surprising number of accounts from ordinary operators — a window importer, a power-washing company, a solo newsletter writer — who had quietly figured out how to make AI dependable.
A pattern kept surfacing. When people get bad results from AI, they almost always reach for a better model or a cleverer prompt. But the research points somewhere else entirely. In a Berkeley study of failed multi-agent runs, expert annotators traced about 42% of failures to specification and design — vague instructions, undefined roles, no stopping conditions — the single largest category, ahead of coordination and verification (MAST). Most AI failures aren’t model problems. They’re management problems.
Three habits do most of the work when you want good results from AI tools, and none of them require code. I think of them as Define, Feed, and Check.
Define: Say what “done” actually means.
Left to guess your unstated requirements, models get them right only about 41% of the time (MAST). And here’s the frustrating part — the AI usually knows your request is ambiguous and answers anyway rather than asking. It behaves like a nervous new hire who won’t raise their hand.
The cheapest fix I found is five words: end your prompt with “ask me clarifying questions first.” It sounds trivial, but forcing that clarify-first step raised one model’s rate of producing working code from about 71% to 81% in a peer-reviewed study (ClarifyGPT, FSE 2024). The refinement that makes it stick: once the AI has interviewed you, have it fold your answers back into a single brief so the definition lives in one place. One person started from a single sentence, let the AI ask 32 questions, and ended up with a spec that surfaced things he’d never considered — a third of the questions caught real gaps (VelvetShark).
Two More Principles
Define with examples, not adjectives. Every strong “make it sound like us” setup I found was built from real samples plus specific vetoes, not descriptors. My favorite framing is the contrast trio — show it a good line, a line that goes too far, and a line that dies of beige — so the model learns the edges of the voice (Search Engine Land).
Ration your rules. More definition is not better definition. Compliance decays predictably as instructions pile up, and experienced practitioners keep their standing instruction files short — often under 60 lines — because models start ignoring bloated rulebooks (HumanLayer). Five clear musts beat twenty nice-to-haves. When a process matters, the move is to write it down as a short standard operating procedure — which is exactly what Amazon’s internal builders landed on after finding that unstructured guidance produced “wildly different results” (AWS).
Feed: Give it the right material, and only the right material.
Whatever you feed the AI, you own. Stanford researchers planted errors in documents they handed to frontier models — wrong dosages, wrong records — and the models abandoned their own correct knowledge and repeated the document’s error more than 60% of the time (ClashEval). The AI trusts your material more than it trusts itself. That’s exactly why feeding it your real memo, transcript, and numbers works — and why those inputs have to be right.
The counterintuitive half is that more context measurably hurts. Chroma tested 18 frontier models across 194,000 calls and every one degraded as the input grew, even on trivial tasks (Context Rot). A short, focused brief beat the same facts buried in a sprawling chat history — roughly 300 words of signal versus 100,000 of noise. The lesson I keep repeating to myself: brief the AI like you’d brief a new hire. Don’t forward the whole thread — write the brief. Anthropic’s own teams frame the goal as finding “the smallest possible set of high-signal tokens” for the job (Anthropic).
For work you do repeatedly, curate a home for it — one focused project per business function, loaded with your handful of best past examples, with every correction folded back in (Amit Kothari). And treat freshness as a chore, not a setup step. Intercom’s teams that hit an 80% AI-resolution rate spend roughly 25 staff-hours a week keeping their knowledge base current (Intercom). The cautionary tale is Air Canada, which was ordered to pay damages after its chatbot cited a stale policy that no longer existed (CBC). Give the documents your AI relies on a shelf life and sweep them periodically.
Check: Let fresh eyes review the work.
The most consistent finding across everything I read is that a reviewer who didn’t see the reasoning that produced the work catches more of its flaws. Asking the same chat to “double-check your work” mostly adds noise; pasting the draft into a fresh chat that doesn’t know who wrote it significantly improves error-catching (controlled experiment, 2026). It makes sense once you name it — asking a model to review its own output is asking the author to review their own book. Models have a documented blind spot for errors in their own work that they’ll happily fix when the same errors arrive as someone else’s (Self-Correction Bench).
A few things sharpen the review. Give the reviewer one lens and a written rubric — grade pass/fail against specific criteria rather than asking “is this good?” (hamel.dev). For high-stakes work, put one seat at the table for a different vendor’s model, because a council of identical models shares identical blind spots (Blattman’s council).
And keep the last gate human. I couldn’t find a single credible account of anyone running production AI work without a person approving the moment something is sent, quoted, or published. My favorite example is a power-washing company whose automation reads the address, checks Street View, and drafts the quote — but a technician taps “approve” before any customer sees a price (n8n) — a blue-collar shop describing quality control better than most software teams do. In everything I read, no one ran production AI work without a person in that last seat.
Quality from discipline
None of this is about having the newest model. Define the work clearly, feed it the right material, and check the output with fresh eyes and a human gate — and ordinary tools start producing dependable results. It’s less like programming a computer and more like managing a talented, eager, slightly overconfident new team member. That’s a skill most leaders already have.
At Atomic Object, this is the same discipline we bring to building software for our clients — clear definition, good inputs, and rigorous review — now pointed at a new kind of collaborator. If you’re trying to get more reliable work out of AI, I’d start with the five-word version: ask it to ask you questions first.