Better AI starts with better judgement 

Bottom line: if you or your team’s AI output is mediocre, it’s most likely that the model isn’t the problem, it’s that your people may not know when to use it, what to expect from it, or which mode to use when they do. 

That’s not a shot at you nor your team. It’s one of the single most useful thing said about AI productivity in months, and it didn’t come from a consultant selling a framework, it came directly from inside Anthropic. 

The comment that started this 

An Anthropic engineer, Thariq Shihipar, recently laid out his own workflow for getting better results out of Claude’s newest models. His core point being that output quality now depends less on the model and more on how well users can spot their own blind spots before they ever write a prompt. He points to techniques like blind-spot passes and structured interviews with the AI as ways to surface what you don’t know before it gets expensive to fix. 

Read that again. The guy building the model is telling you the bottleneck is you. Is that uncomfortable? For some, it may be, but for those who truly want strong integration into their tool stack, it should be affirming. It’s a concept that applies to both the technological world and most other tools, and it applies to every knowledge worker prompting an AI model today, on any platform. 

It’s a requirements problem, not a tool problem 

Here’s what we’re seeing across every engagement: Leaders want to realize and maximize their business and believe AI and automation with AI will get them there. Concurrently (or in parallel), teams jump straight to prompting before they’ve diagnosed the gap they’re actually trying to close. They ask the model to fill in blanks they haven’t identified themselves. Then the output comes back generic, or wrong, or “not quite it” — and the conclusion is “the AI isn’t good enough” or “AI doesn’t understand what I need from it”.  

Wrong conclusion. The model did exactly what it was asked. Nobody defined the requirement. As we’ve mentioned before, AI is not the “magic answer box”, it maps your request to what it believes is the most likely response. Incredibly efficient, but akin to having the apprentice handing you tools that it *thinks* you need without full understanding of your desired outcome or solution.  

This is a maturity marker. Teams that move from “does the model work?” to “do I actually understand what I’m asking for?” get dramatically better outputs, immediately, with the same tool and the same subscription. No model upgrade required. That’s the signal leadership should care about, it’s cheaper and faster than any procurement decision. 

Three questions your team should be answering before they prompt 

The first thing we see, once we get teams over the hump of understanding the basics of prompting, is they then move to syntax and prompt engineering (how to phrase a prompt). That’s not wrong; it’s just incomplete. Phrasing is the last five percent. The three questions below are the other ninety-five, and skipping them is why “prompt training” so often doesn’t move the needle: 

1. When should I even be doing this with AI? Not every task is an AI task. Some things are faster done by hand, some require judgment AI shouldn’t own yet, and some are exactly the kind of repetitive, well-bounded work AI eats for lunch. Knowing which bucket you’re in is step one, and most people skip it entirely. 

2. What outcome should I actually expect? A first draft is not a final answer. A brainstorm is not a decision. Setting the wrong expectation is why people either over-trust AI output (ship it as-is, get burned) or under-trust it (dismiss a genuinely useful draft because it wasn’t perfect on the first pass). And it goes further: people often take the first response as the final answer and don’t probe, don’t verify, and don’t align. It is imperative that your team not only trust, but verify, but that they also apply their understanding of the response. This could manifest itself as additional, probative questioning about the response and/or challenging it outright if something looks off. They key in so doing, however, is to not get stuck and lose any accumulated productivity gains.  

3. Which mode fits the job? Think of it like a toolbox. A screwdriver, a power drill, a contractor crew, and an assembly line all drive a screw — but you’d never use them interchangeably. Nobody hires a crew to hang one picture frame, and nobody hand-turns a screwdriver 500 times on a production line. Most AI users own exactly one tool — the simple prompt — and reach for it regardless of the job. This is where most people are flying blind: 

• Simple prompt (the screwdriver): One question, one answer, done. Good for lookups, quick rewrites, low-stakes drafts. Fast, cheap, and useless the moment the job gets bigger than one turn. 

• Single-shot vs. multi-shot (hand tool vs. power tool): Do you need one clean output, or a back-and-forth where the model asks you questions and narrows in? Complex or ambiguous work almost always needs multi-shot. People default to single-shot because it feels faster, then wonder why the result missed the mark — that’s hand-turning a screw that needed torque. 

• Agent (the contractor crew): The task requires multiple steps, tool use, or ongoing autonomy (research, multi-file work, monitoring). Using a simple prompt for an agent-shaped problem is like emailing someone one instruction at a time instead of handing them the project and letting them run it. 

• Automation (the assembly line): The task is repeatable and well-defined enough that a human shouldn’t be re-prompting it every time. If you’re running the same multi-shot conversation weekly, it should be a pipeline, not a chat window. 

Get the when, what outcome, and which tool right and the phrasing of the prompt itself stops being the thing that matters most. That’s the point most training programs miss: they hand people a better screwdriver and wonder why the assembly-line jobs still aren’t getting done. 

Where this matters for your business 

This is the exact gap LumenForge’s Fluency Competency Model was built to close inside the LumenForge People Fluency Framework. The issue is not just “how to write a better prompt.” That is a tactic, and tactics age out fast as models change. The durable skill is knowing when AI belongs in the work, how to frame the task, what level of confidence the output deserves, and where human judgment must stay in the loop. 

That is why Prompt Discipline is a practical entry point. It is a focused one-hour workshop that gives teams a shared operating language for using AI well. We cover the basics of prompting, including single-shot, multi-shot, and iterative prompting, but the larger point is discipline: grounding the model with the right context, defining the outcome before asking for output, checking for gaps, and knowing when to stop, verify, or escalate. 

The key framing is simple: AI fluency is not just knowing how to prompt. It is knowing when to use AI, how to engage it, and when not to over-rely on it. That “when and how” distinction matters because most AI failures in everyday work do not come from bad syntax. They come from unclear requirements, missing context, weak verification, and treating a first response like a finished answer. 

The Fluency Competency Model turns that into a maturity path. It helps leaders see whether their teams are experimenting casually, using AI with basic prompting skill, applying grounded and iterative workflows, or operating with the judgment, governance, and repeatability needed for real business value. Prompt Discipline sits at the front end of that path. It gives people enough practical fluency to stop using AI like a novelty and start using it like a work system. 

The LumenForge People Fluency Framework then connects that individual skill to the broader operating model: people, process, governance, and value. That is the part most organizations miss. They train people on prompts, but they do not build the surrounding muscle: shared expectations, responsible use, repeatable patterns, human oversight, and decision discipline. Without that, AI adoption stays fragmented, inconsistent, and hard to scale. 

Net: your AI ROI ceiling isn’t set by the model you license. It’s set by whether your people know what they don’t know before they ask. Model upgrades are the easy button. Requirements discipline is the actual lever and it’s the one most organizations haven’t pulled yet. 

If you want a clear-eyed read on where your team’s AI fluency actually stands — not where you assume it is — the AI Agent Exposure Check is a free 10-minute starting point (https://www.lumenforge.ai/ai-agent-exposure-check/). And if you want to go even further on a journey with integrating AI into your environment, set up your free AI Reality check at https://bookings.cloud.microsoft/book/LumenForgeAdvisors@lumenforge.ai/?ismsaljsauthenabledand. 

References 

The-Decoder. (2026, July). Anthropic developer shares prompting tips for Fable 5 that focus on finding your own blind spots first. Retrieved from https://the-decoder.com/anthropic-developer-shares-prompting-tips-for-fable-5-that-focus-on-finding-your-own-blind-spots-first/