The Input Layer

Prompt Engineering Isn’t Dead. You’re Just Doing It Wrong.

My best prompts are around two thousand words long.

That fact tends to stop conversations. People who’ve been told that prompt engineering is dead, that models are smart enough now to figure out what you want from a few sentences, that the whole concept has been “absorbed” into normal AI usage, don’t know what to do with the idea that someone sits down and writes a prompt the length of a long-winded wedding toast. It doesn’t fit their model of how this works. It suggests that the person typing is doing something categorically different from what they’re doing, and that what they’ve been calling "prompting" might not be the same activity at all.

It isn’t.

A Microsoft survey of 31,000 workers recently ranked “prompt engineer” second to last among roles companies plan to add.[1] Fast Company reported in 2025 that 68% of firms treat it as basic training, not a specialty.[2] The cottage industry of prompt-engineering courses and certifications has mostly collapsed. And the consensus, at least among people who write about AI for a living, is that the skill has been automated away by smarter models.

It’s true that some of the old tricks don’t matter as much anymore. Modern AI reasoning models don’t need me to preface every prompt with “you are a world-class expert in securities law” (they’re already pretty damn good at reasoning about securities law). But they still can’t read my mind. A model that can think doesn’t automatically know what to think about. It doesn’t automatically know my client’s risk tolerance, my counterparty’s negotiating posture, or which of the fourteen issues in this contract are the three that actually matter. I still have to tell it all of that. That’s the part of “prompt engineering” that didn’t die: the ability to communicate with AI the way a senior professional communicates with a talented but dangerously literal colleague (in detail, with precision, leaving no room for ambiguity about the former’s expectations of the latter). That ability is currently my greatest professional edge, and is becoming the most important skill in the white collar economy. Unfortunately, almost nobody understands it or is teaching it correctly.

I run an AI-native law firm. I draft contracts, review opposing counsel’s redlines, analyze regulatory questions, negotiate deal documents, and handle client communications. AI is my primary collaborator on all of it. After a couple of years at this full-time, I can tell you that the quality of what comes out is almost entirely a function of what I put in. The magic lives in the input layer.

The Genie

The best mental model for using AI effectively is one everybody already knows and yet almost nobody actually applies: the genie.

You get three wishes. The power is functionally infinite. And yet pretty much every genie story ever told runs on the same premise: the genie does exactly what you say, not what you mean. Wish for a million dollars and it falls from the sky and flattens your house. Wish to be the most powerful person alive and wake up alone on a dead planet. The genie isn’t cruel, it’s just literal, so the entire drama lives in the gap between what you said and what you thought you said.[3]

Large language models are functionally very literal genies. There’s a mechanical reason why vague prompts produce bad output that goes beyond the metaphor. These models were trained on the corpus of essentially the entire internet. When you give one a vague instruction, it does what it was trained to do: it regresses to the mean of everything it’s ever seen. In other words, it gives you the average of what the internet would say (with extra emphasis on Reddit posts, em-dashes, and “it’s not this, it’s that” prose structure). You know what that looks like. Everyone does by now. It’s the flat, confident, vaguely competent text that has become its own aesthetic category: AI slop. You recognize it instantly in other people’s LinkedIn posts.

But when you feed the model instructions that are specific and detailed enough to pin it to one narrow path, something different happens. The output snaps into focus. It stops reaching for the average and starts executing on the particular. The key insight, the one that took me the longest to learn, is that your instructions need to do two things: tell the model what to produce, and close off every other thing it might produce instead. It’s not enough to describe what you want. You need to describe what you want precisely enough that there is no room left for the model to wander back toward the generic. You are not just pointing at a destination. You are building a corridor. (See what I did there?)

In classic genie stories, the danger is never that the genie is weak, it’s that the wisher is lazy. They want the result without doing the cognitive work of figuring out what, specifically, they actually want. And this maps exactly onto how most people use AI. You have to be able to see the output clearly in your own mind before you can describe it to the model. Any fuzziness in your thinking, any ambiguity about what you’re actually looking for, gets faithfully reproduced as fuzziness in the output.

The good news is that what I’m describing is a mindset, not a technical skill. I want to be clear about this because it matters. I am genuinely bad at learning software tools. I have ADD. My entire life, organizational systems have defeated me. Clients hand me Asana boards and Jira tickets and Notion dashboards and other similarly horrible things, and I struggle with all of them. The dashboards, the feature menus, the nested settings panels just don’t stick for me. But AI doesn’t work like that. The interface is just a text box. The input method is plain English. You type or you speak. There is no feature to find, no workflow to memorize, no certification to earn. The only thing you need is the discipline to translate everything in your head, all of the context, the intent, the constraints, the judgment calls, into language precise enough that a literal-minded machine can act on it without guessing. That’s the whole skill. It’s the skill of being clear about what you want. And for some reason, almost nobody is teaching it that way.

The Abdication Trap

In January 2024, an Illinois woman named Graciela Dela Torre settled a disability claim and signed a full legal release. The case was dismissed with prejudice, meaning it was over. A year later, unhappy with the settlement, she asked her attorney to reopen it. He told her the release was enforceable. She uploaded his response to ChatGPT and asked if she was being gaslit. ChatGPT said yes. She fired her lawyer. Over the next several months, with ChatGPT as her legal advisor, Dela Torre filed more than 60 documents across two federal cases, nearly all drafted by the chatbot, many citing cases that don’t exist. Her former counterparty spent $300,000 defending against an avalanche of fabricated filings in a case that had already been resolved. This month, they sued OpenAI for $10.3 million.[4]

Dela Torre isn’t a lawyer. But the dynamic that destroyed her case is now playing out across the profession at scale. More than 550 documented instances of AI-hallucinated citations have surfaced in legal filings nationwide. The rate went from two per week in early 2025 to two or three per day by December.[5] An Am Law 100 firm has been caught multiple times.[6] A New York judge found violations of four Rules of Professional Conduct and wrote that lawyers “may not abdicate these core professional responsibilities to an AI platform.”[7] Every one of these cases follows the same script: someone asked a chatbot to do their thinking for them, took whatever came back, and acted on it without adding anything of their own.

The lesson most people draw from these stories is that AI is dangerous for professional work. Ezra Klein made a version of this argument last year: that AI is “fundamentally problematic” for serious writing because it automates the part of the process where the thinking actually happens, the struggle of figuring out what you believe, the hours spent wrestling a messy idea into a clear sentence.[8] Skip that struggle and you’ve skipped the only part that mattered.

I think Klein is right about the diagnosis and wrong about the conclusion. The cognitive work is the work—no argument from me there. But Klein assumes that if AI is involved, the struggle got skipped. There’s another possibility that he doesn’t consider: you do the hard thinking first, all of it, the planning, the analysis, the strategy, the judgment, and then you bring in the AI to execute on that thinking, or to pressure-test it, or to help you see what you missed. Every article I publish draws replies from people breathlessly insisting I used AI to write it. They’re right. I did. I’m using it right now. What they’re wrong about is the assumption that the AI did my thinking for me. It didn’t.

That’s the category error at the center of this whole discourse. Everyone draws the line between “human-made” and “AI-made,” as if those are the only two categories. The real line is between work where the human did the cognitive labor and work where they didn’t. On one side is slop: someone feeding a one-sentence prompt into a chatbot and shipping whatever comes back (the hallucinated briefs, the LinkedIn posts that read like they were generated in a microwave, every piece of writing where you can feel that nobody was home when it was produced). On the other side: work where a human showed up with something worth saying, and used AI to say it better, faster, or at a scale they couldn’t reach alone. The presence of AI in the process tells you nothing. The presence of a thinking human tells you everything.

And the temptation to inhabit the wrong side of this divide is getting stronger as models improve. The better AI gets at producing polished, confident, professional-looking output, the easier it becomes to trust the surface, to skip the verification, to let the packaging convince you. The quality of the wrapping is rising just fast enough to make the shortcut feel safe. But the underlying failure modes that are inherent to the technology of LLMs haven’t changed, and the polish is a trap.

AI does not eliminate the need for careful thought. You have to do the analysis before you open the chat window. You have to know what you want, why, under what constraints, and where the traps are. The prompt is the work product of your thinking. It is not a substitute for your thinking. Professionals who internalize this will be fine. The ones who treat AI as a way to skip the hard parts will get caught, sanctioned, or replaced. Some will manage all three.

The Output Fallacy

If the abdication trap is the individual failure mode, the output fallacy is the institutional one. It is the belief, shared by most of the legal tech industry and a depressingly large portion of LinkedIn, that the value of AI lives in its outputs and can be captured by optimizing them.

There are two versions of this mistake. The consumer version is the “7 Amazing Prompts That Will Transform Your Practice” content that clogs every professional’s social media feed. These posts promise that some template, some incantation, will reliably produce great results across contexts. I promise you it won’t. A prompt that works for reviewing a Series A preferred stock purchase agreement will crater when you point it at a commercial lease. The context is different. The risks are different. The client cares about different things. The entire analytical frame is different. Good prompts don’t generalize, because the judgment baked into a good prompt is specific to the situation. The people selling universal prompts are selling something that stops working the moment you need it for anything real.

The enterprise version of the mistake is similar, but vastly more expensive on a per-seat basis: vertical-specific (e.g. legal) AI wrappers. Companies like Harvey, now valued at $11 billion,[9] have raised staggering sums on the premise that fine-tuning large language models on specific legal corpora (SEC filings, litigation briefs, particular firms’ contract templates) will produce better results than a frontier model in the hands of a skilled operator. This is the output layer.

But the output layer isn’t where the bottleneck sits. Every frontier model has already ingested the whole internet. Every publicly available legal template, every SEC filing on EDGAR, every published opinion, every law review article. Claude knows what a good merger agreement looks like. So does ChatGPT, Gemini, Grok, DeepSeek. The model’s legal knowledge was never the constraint. The constraint was always the human prompter’s ability to say, with precision, what this agreement needs to do for this client in this deal.

This is true of individual prompts, where the value is contextual specificity. It’s even more true of the reusable skill systems I’ve built on top of them (more on this shortly). I’ve actually tested this from the other direction. I’ve had people try to reverse-engineer my Claude skills by studying my outputs, using AI to analyze what I produce and reconstruct the instructions that generated it. They never get close. They can mimic stylistic patterns, approximate my voice. But what my skills actually contain is not a description of what the output should look like. It’s a detailed operating procedure for how the output gets created: decision trees, analytical frameworks, sequencing logic, edge-case handling, judgment calls about when to be aggressive and when to hold back. You can’t see any of that by studying the finished product. You might be able to identify every ingredient in a chef’s dish by tasting it, but you’ll still have no idea how to make it, because the recipe isn’t just in the ingredients, it’s in the procedure (chopping, emulsifying, letting it sit in the fridge overnight). Same principle here. A finished contract shows you what a great lawyer decided. It doesn’t show you how she decided it, what she considered and rejected, or the order in which she worked through the issues. The process is invisible in the product. My skills encode the process.

This is exactly the error the legal tech industry is making at scale. You don’t get at what makes a great contract lawyer great by training on thousands of that lawyer’s (or firm’s) contracts. That’s pattern-matching the output layer. What makes them great is the process: the sequence of judgment calls they make while drafting, the analytical moves they apply to each clause, the instincts about when to push and when to concede. That process lives entirely in the input layer. It has to be individually encoded by someone who understands both the domain and how to talk to the model. No amount of post-training on finished documents will get you there.

The entire legal tech industry is optimizing the output layer. The value lives in the input layer. They are building elaborate, expensive infrastructure around the wrong end of the pipe. And charging a thousand dollars per lawyer per month for it.

Talk to It Like a Partner

So what does the input layer actually look like?

It looks like a conversation. Not a search query. Not a task list. It looks like the conversation a senior partner has across the desk with an associate she’s handing a complex assignment to. The kind of briefing where you explain why the client cares about this deal, what the counterparty is likely to push back on, which provisions are dealbreakers and which are chips to trade, what tone to strike, how aggressive to be, and the six ways this particular assignment could quietly go sideways if the associate isn’t paying attention.

People are surprised my prompts run to two thousand words. They shouldn’t be. Think about how much context a senior lawyer conveys when handing off a deal. All the "unstated assumptions" that should likely be stated for good measure. The relationship dynamics. The judgment calls about what matters and what’s noise. A typical prompt of mine covers the client’s business model and risk tolerance, the history with this particular counterparty, which deviations from market terms are acceptable, how the comments should read stylistically, what the client’s other outside counsel has historically fixated on. That’s before I get to the specific legal instructions about the document itself.

Now picture the associate you’re briefing: brilliant, has read every contract ever published, works at inhuman speed, but takes every instruction with complete literalism and has zero independent judgment about what you probably meant. You can’t gesture. You can’t hint. The genie needs the complete wish. And the wish has to be precise enough that the output matches the picture you already have in your mind. If you can’t clearly see what you’re looking for, you will not be able to describe it, and the model will turn your vagueness into mush.

This is where my background turns out to be unexpectedly useful. I studied philosophy at Williams and Oxford. I’m a devotee of Wittgenstein and Ryle, of the “ordinary language” philosophical tradition, which is obsessively concerned with one question: what do we actually mean when we say something? What work is a concept doing in a sentence? Where does ambiguity hide in apparently clear language?[10] My professional life is drafting commercial contracts, which is the same exercise applied to business relationships: close every gap, capture every edge case, make sure each term does exactly what it’s supposed to do and nothing else. It turns out that writing a good contract and writing a good prompt involve the same cognitive act. In both cases, you are trying to eliminate every space where a literal interpreter could go wrong.

Lawyers should be great at this. The skill set of giving a clear, detailed, context-rich assignment to a junior lawyer maps almost perfectly onto prompting. Same vocabulary. Same analytical instincts. Same habit of anticipating where things could go wrong. The irony is hard to overstate: the profession best equipped to use AI well has been the slowest to try, and when it tries, it reaches for a thousand-dollar-a-month wrapper instead of using the ability it already has.

The Compounding Flywheel

Everything I’ve described so far is about the single prompt: being detailed, being specific, embedding your judgment into the instructions. That’s the first half of the equation, and it’s where even the most effective users of AI tend to stop. But a great prompt is still a one-time event. It solves one problem. The real unlock, the one most people miss entirely, is what happens when you start encoding your prompting patterns into reusable systems and then refining those systems against the results of real work.

A prompt is a set of context-specific instructions: here is this client, this deal, this document, here is what I need. A skill is a set of procedures: when you receive a contract for review, here is the sequence of analysis you perform, here are the decision rules for each type of provision, here is how you weigh competing risks, here is what the output should look like when you’re done. The prompt encodes context. The skill encodes process. And process is what compounds efficiency.

The mechanism works like reinforcement learning applied to your own practice. Every time AI output surprises you, good or bad, stop and ask why. Contract review catches a subtle interaction between two clauses you hadn’t thought to flag? Figure out what in your instructions led there. Encode it in the skill file. Draft comes back with a tone problem, misses a standard risk, structures an argument wrong? Figure out what was missing. Update the skill. Each cycle makes the next output incrementally better. Your judgment accretes in the system like sediment building a riverbed.

Do this for months, across dozens of matters, and the skill becomes a living document encoding hundreds of micro-decisions about how you practice. It knows that on technology licenses, you always flag IP assignment provisions lacking carve-outs for pre-existing IP. It knows your client emails never open with “I hope this finds you well.” It knows you want surgical redline edits with explanatory comments, not wholesale rewrites. It knows how to mark up a first draft from a sophisticated counterparty differently from a template sent by a first-time founder. None of this came from the base model. All of it came from paying attention, one matter at a time.

After enough iterations, the practical result looks like a magic trick. I can upload a contract from opposing counsel, type “plz fix,” and get back a perfectly redlined document with tracked changes, explanatory comments, and risk analysis. Formatted the way I’d format it. Catching the issues I’d catch. In a voice that reads like my own work product. Not because the AI woke up one day with brilliant instincts about contract review. Because I spent hundreds of hours teaching it how I think.

The skill file that enables that output is valuable intellectual property. It contains something no generic training corpus has and no wrapper can replicate: the specific professional judgment of a specific practitioner, refined across real client work. My judgment, bottled. And it gets better every week.

The New Bottleneck

If the input layer is where the value lives, the question every organization should be asking is: who are the people who are good at this, and what makes them good?

The answer is not what most firms expect. The people who will be best at communicating with AI are not, primarily, technologists. They’re not the people who understand transformer architectures or can fine-tune a model. They’re domain experts with a specific and somewhat unusual combination of traits: deep knowledge of their field, the ability to articulate that knowledge with extreme precision, and the discipline to treat every AI interaction as an opportunity to encode their judgment a little further. They are, in short, the people who were already the best at their jobs, plus this one new dimension.

Think about what a 2,000-word prompt actually requires. You need to know your domain deeply enough to identify the issues that matter. You need to understand your client’s situation specifically enough to set the right priorities. You need the linguistic precision to describe all of this without leaving gaps for a literal machine to misinterpret. And then, at the skill level, you need the metacognitive discipline to watch the output, diagnose what’s working and what isn’t, and iteratively refine the system. That combination doesn’t show up on any standard hiring rubric. But it’s the combination that produces 10x output.

The World Economic Forum projects that 39% of core professional skills will be transformed by 2030.[11] Oxford Internet Institute research found that professionals with AI capabilities earn 21% more than comparable peers. Multiple AI competencies push that premium to 43%.[12] The market is beginning to price this in. But most organizations still don’t understand what they’re selecting for. They post job listings asking for “AI literacy,” which is roughly like asking for “computer literacy” in 1995. They buy platform licenses and assume the technology will do the work. They send associates to prompt-engineering workshops that teach the parlor-trick version. None of this produces the capability I’m describing.

What produces it is a person who has done the hard work of building the skills, who has iterated across hundreds of real matters, who has developed an intuition for how to talk to a system that is powerful and literal and surprisingly responsive to the right kind of instruction. That person, equipped with their compounding skill library, can do in an afternoon what used to take a team a week. Not because the AI is magic. Because the human operating it spent two years learning exactly how to wish. That is a capability firms and clients should be willing to pay enormously for, and they will, once they understand what it actually is.

Behind the Curtain

I’m not giving away the full playbook. And if you’d like access to the markdown files that hold my skills, I’d be happy to entertain offers starting in the millions of dollars (please send me a DM). But one technique is worth sharing, because it points at something most people don’t realize about the frontier of this work.

Many of my skills end with a version of this line: “Before delivering this work product, please consider carefully whether, if I hand this to my client as-is, there is anything in here that would embarrass me.”

That is not the same as writing “check citations” and “verify formatting” and “make sure the edits are consistent,” even though it accomplishes all of those things. It’s better. And explaining why requires saying something that sounds strange but is true: the most advanced AI models have something that functions like a personality, and you can prompt it.

Claude Opus 4.6, the model I work with most, doesn’t just execute instructions in sequence. It reasons about them. It has an internal monologue. When I end a skill with “would this embarrass me in front of my client?,” the model’s reasoning doesn’t just produce a checklist, but something more like conscience. It starts thinking about my reputation, my client relationship, what it would mean for a sloppy citation or a tonal mismatch to slip through. The citations get verified, the formatting gets cleaned, the legal edits get cross-checked for consistency. But not because I itemized each task. Because I gave it something closer to a motive. The frame “don’t embarrass me” activates a deeper layer of care than any set of specific instructions I’ve been able to write. I’ve tested this head-to-head, many times. The version with the frame wins.

I know how that sounds. I genuinely don’t care, because the difference in output quality is real, repeatable, and large. If you’ve spent serious time working at the frontier of these systems, you already know what I’m describing. If you haven’t, I understand the skepticism. But this is exactly the kind of counterintuitive discovery that separates someone who uses AI casually from someone who has gone deep enough to develop a feel for how it thinks. No manual teaches this. No wrapper builds it. You find it by paying close attention, over many, many hours, to what actually works even if it seems like it shouldn’t.

There are dozens of techniques that I’ve found just like this one. Each represents a hard-won discovery about how to coax the best work from a system that is immensely powerful, frustratingly literal, and responsive to kinds of instruction most people would never think to try. Together, they’re the reason I can do things with AI that most professionals can’t. Not because I’m a better technologist. I’m probably worse at technology than half the people reading this. It’s because I put in the hours on the input layer, and the input layer is where the leverage lives.

The Input Layer

Prompt engineering isn’t dead. The trivial version died, the version that was always just a parlor trick. What survived is the discipline of turning professional judgment into precise instructions, then transmuting those instructions over time into reusable intelligence, compounding in the reinforcement learning environment of the white collar workplace, that becomes, in every meaningful sense, valuable IP.

The people who have “tried” AI and declared it overhyped are typing three sentences into a chatbot and blaming the model when the output is flat. The legal tech companies building wrappers around the output layer are solving a problem that frontier models solved two generations ago. The LinkedIn influencers promising seven prompts that will change your life are offering fortune cookies to people who need a meal.

The value was always in the input layer. The professionals who build their work around that insight, who treat every interaction as a chance to encode their judgment a little more precisely and compound their effectiveness a little further, will produce work that looks like sorcery to everyone still rubbing the lamp and mumbling.

The genie is here. Learn to wish.

Notes

1. Microsoft-commissioned survey of 31,000 workers across 31 countries: Prompt Engineer ranked second to last among new roles companies plan to add. Salesforce Ben (May 2025).

2. Fast Company reported (May 2025) that 68% of firms now provide prompt engineering as standard training rather than a standalone role.

3. Nick Bostrom’s “paperclip maximizer” thought experiment, introduced in Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014), illustrates the same structural problem at civilizational scale: an AI given a literal objective (maximize paperclip production) and insufficient constraints on how to achieve it will pursue that objective in ways its designers never intended, including converting all available matter into paperclips. The gap between instruction and intent is not merely a prompting inconvenience. It is the central problem of AI alignment.

4. Nippon Life Insurance Co. of America v. OpenAI, Inc. (N.D. Ill., filed March 5, 2026): Nippon alleges ChatGPT encouraged former disability claimant Graciela Dela Torre to breach a settlement, fire her attorney, and file 60+ AI-generated court documents including fabricated citations (e.g., Carr v. Gateway, Inc., which does not exist). Nippon claims $300,000 in defense costs. Seeks $10.3 million including punitive damages. OpenAI called the complaint meritless. The Hill, ABA Journal, Law360, IBTimes, FindLaw (March 2026).

5. AI hallucination cases in legal filings: As of early 2026, publicly reported cases exceeded 550 nationwide. Damien Charlotin’s tracking database, featured in the LA Times, Volokh Conspiracy, and 404 Media, documents instances worldwide. Bloomberg Law reported the rate increased from approximately two per week in early 2025 to two to six per day by December 2025.

6. Gordon Rees Scully Mansukhani (Am Law 100): Multiple reprimands. Jackson Hosp. & Clinic Inc.: “pervasive inaccurate, misleading, and fabricated citations.” Second reprimand in Villalovos-Gutierrez v. Pol (December 2025). Above the Law (February 2026).

7. Cassata v. Michael Macrina Architect, P.C. (Suffolk County, N.Y. 2026): Justice Linda Kevins found defense counsel violated Rules 1.1, 1.3, and 3.1 for submitting AI-generated filings without meaningful review, cautioning that lawyers “may not abdicate these core professional responsibilities to an AI platform.” New York Commercial Division Practice (February 2026).

8. Ezra Klein, “The Case Against Writing With AI,” interview with David Perell on the How I Write podcast (May 28, 2025). Klein argued that AI automates the wrong part of the writing process, that the cognitive struggle of composing is where the thinking happens, and that outsourcing that struggle to AI means outsourcing the thinking itself.

9. Harvey AI valuation: $3B (Feb. 2025, Sequoia-led Series D), $5B (June 2025, Kleiner Perkins/Coatue Series E), $8B (Dec. 2025, a16z), reportedly $11B (Feb. 2026, $200M round led by Sequoia and GIC). Total capital raised exceeds $1.2B. ARR reached $190M by end of 2025, nearly doubling from $100M five months earlier. Pricing ~$1,000–1,200/lawyer/month. Forbes, TechCrunch, Law.com (Feb. 2026). Reddit "ChatGPT wrapper" discourse: LawSites (Sept. 2025).

10. Ludwig Wittgenstein, Philosophical Investigations (1953); Gilbert Ryle, The Concept of Mind (1949). The ordinary language tradition holds that philosophical problems often arise from misunderstandings of how language actually works, and that careful attention to the use and meaning of words dissolves many apparent confusions.

11. World Economic Forum, Future of Jobs Report 2025: 39% of core skills expected to change by 2030. AI/big data top the fastest-growing competency list. 77% of employers plan reskilling. Skills demand changing 66% faster in AI-exposed occupations.

12. Oxford Internet Institute: AI-skilled professionals earn 21% more than peers; multiple AI competencies push the premium to 43%. The Interview Guys (Feb. 2026). Dice 2025 Tech Salary Report: AI/ML skills carry 17.7–18% premium.