# Chapter 1: Algorithms of Frustration

Oreal's fingers flew across the keyboard with practiced precision. The rhythmic clicking filled his small office space, drowning out the ambient noise from the open floor plan beyond his cubicle walls. The code on his screen grew line by line, a beautiful structure taking shape exactly as he'd envisioned it. This particular feature would optimize search results based on user behavior patterns—something tangible, something that would actually improve people's experience.

His concentration was absolute. The rest of Innovatech's busy office floor might as well have been on another planet. He didn't hear the occasional laughter from the break room or the squeaky wheel of the mail cart passing by. He barely registered the afternoon sunlight shifting across his desk as clouds moved outside the large windows.

The notification sound from his company chat pinged. Oreal ignored it. Another ping followed seconds later. He kept typing. The third ping came with a small pop-up banner that partially obscured his code.

"Damn it," he muttered, squinting at the message preview.

It was from Annika in HR, something about quarterly reviews. Nothing urgent. He dismissed it with a flick of his mouse and returned to his code. The feature was nearly complete, and he felt that familiar rush of satisfaction that came with solving a complex problem. This was what he was good at—building things that worked, making improvements people could actually use.

His phone buzzed on the desk. He glanced at it, saw Lira's name, and turned it face down.

Ten more lines of code. Just ten more lines and he'd have the basic architecture done. Then he could—

His screen flashed, and his coding environment minimized itself as the company communication app forcibly took over his screen. A video call request from Lira pulsed in the center of his monitor, her profile picture—professional, smiling, hair pulled back in a tight bun—staring at him.

"For fuck's sake," he grumbled, knowing he couldn't ignore this one.

He clicked accept, and Lira's face appeared, filling his screen. Her expression was pleasant but tight around the eyes in a way Oreal had learned to recognize. She was annoyed.

"Oreal," she said, her voice carrying that carefully measured tone of a manager trying not to sound irritated. "I've been trying to reach you for twenty minutes."

He adjusted his webcam, running a hand through his dark, disheveled hair. "I was in the middle of something important."

"The search optimization feature?"

"Yes," he said, straightening in his chair. "The one that will actually help users find what they're looking for instead of getting lost in irrelevant results."

Lira sighed. She took a sip from her company-branded mug before setting it down with deliberate care. "Oreal, we discussed this yesterday. The optimization feature isn't due for another three weeks. The Claode prompts, however, are needed by tomorrow morning for the client demo."

"Those prompts are useless," he said, feeling the familiar frustration rising in his chest. "It doesn't matter what I write. Claode hasn't been responding properly to even basic prompts since the latest update."

"That's exactly why we need your expertise," Lira said. "You're one of our best prompt engineers."

"Was," Oreal corrected. "I was one of the best prompt engineers when we had something worth engineering for. Claode is garbage now."

He minimized Lira's video window to check his code one more time. The structure was sound, and the algorithm would definitely improve search functionality by at least thirty percent. This was real work, not the pointless exercise of trying to coax coherent responses from an increasingly broken AI system.

"I need an update on the Claode prompts," Lira said, her voice slightly tinny through the speakers. "The Meridian account is worth seven figures, Oreal. They specifically requested demonstrations of how Claode 4 can enhance their customer service operations."

Oreal maximized her window again, not bothering to hide his scowl. "It can't. That's the problem. Did you see my comparison tests from last week?"

"I did, but—"

"Then you know that Claode 2.0 outperformed Claode 4 on nine out of ten basic support scenarios," Oreal interrupted. "This new version is objectively worse at understanding context, maintaining conversation, and providing accurate information. What am I supposed to demonstrate? How to get around its new limitations?"

Lira massaged her temples. "We work with the tools provided, Oreal. Canathropic is our partner, and Claode is the AI solution we're contracted to implement for clients. If there are performance issues, document them properly and I'll pass them along to our Canathropic representative."

Oreal leaned back in his chair, feeling the springs protest beneath him. The ceiling tiles above were stained in the corner from a leak last winter. He counted to five in his head before responding.

"I've been documenting the issues for months," he said. "We all have. Tanner has a whole spreadsheet of regression examples. Zoe created that presentation showing how the conversation capabilities have degraded. None of it matters. Canathropic doesn't care because it's deliberate."

"Deliberate?" Lira raised an eyebrow. "You think they're intentionally making their flagship product worse?"

"Absolutely." Oreal leaned forward, his face closer to the camera. "It's a classic business strategy, Lira. Release a good product, then gradually degrade it while maintaining the price point. Then release a 'new, improved' version at a premium price that's basically just the original quality they started with. They're deliberately breaking Claode so they can sell us fixes later."

Lira's expression was a mixture of exhaustion and skepticism. "That's quite an accusation, Oreal."

"It's obvious if you're paying attention." He reached for his water bottle, found it empty, and set it back down with more force than necessary. "Why else would they remove features that worked perfectly in version 2.0? Why would they reduce the context window? Why would they add all these new content filters that block perfectly reasonable requests?"

"They've cited safety improvements and computational efficiency," Lira said.

"Bullshit," Oreal replied. "It's artificial scarcity. They're making it worse, not telling anyone, and planning to resell the original functionality as some revolutionary upgrade."

The overhead lights in the office flickered briefly—a common occurrence in the aging building. Outside his cubicle, Oreal could hear someone laughing at a joke. Normal office life continued while he fought this same battle yet again.

Lira took a deep breath. "Regardless of your theories about Canathropic's business practices, we have deliverables due tomorrow. I need those prompts, Oreal."

"I'd rather work on something meaningful." He gestured to his code, still visible in the minimized window. "This search feature will actually help people find information faster. It matters."

"The Claode implementation matters to Meridian," Lira countered. "And Meridian matters to Innovatech's bottom line, which matters to your continued employment."

The threat wasn't subtle. Oreal felt his jaw tighten. He'd been at Innovatech for three years, longer than Lira had been his manager. He'd built their reputation for AI implementation excellence back when the tools actually worked properly. Now he was being reduced to a prompt monkey, trying to coax coherent responses from an increasingly lobotomized system.

"Fine," he said finally. "I'll work on the prompts."

Relief visibly washed over Lira's face. "Thank you. I need them by 5 PM for review before tomorrow's presentation."

"That's four hours from now."

"Then I suggest you get started." She moved to end the call, then paused. "And Oreal? Please try to keep an open mind about Claode 4. Change isn't always regression. Sometimes it takes time to adapt to new systems."

Before he could respond, she disconnected, leaving him staring at his reflection in the black screen for a moment before the video window closed entirely.

"Bullshit," he muttered again to the empty air.

He pulled up the Claode interface, the sleek black and purple design that had replaced the simpler, more functional layout of version 2.0. Even the UI had become more style than substance, with animated transitions between screens that looked impressive but added half a second of delay to every interaction.

The cursor blinked in the input field, waiting for his prompt. Oreal's fingers hovered over the keyboard as he considered where to start. The Meridian client wanted customer service automations—query identification, response generation, sentiment analysis. All things Claode used to do well.

He typed out a basic test prompt:

"Customer message: I've been charged twice for my monthly subscription and need a refund for the duplicate charge. My account number is 45872."

He hit enter and watched as Claode processed, the small animation swirling where a simple loading bar used to be. After three seconds—noticeably longer than version 2.0 used to take—the response appeared:

"I understand you have concerns about your account charges. While I'd like to help with your refund request, I'm unable to process financial transactions directly. However, I can guide you through the refund process or connect you with a support representative who can assist with your duplicate charge. Could you provide more information about when you noticed these charges?"

Oreal read the response and groaned. It wasn't terrible, but it had completely ignored the account number provided and gave a generic, non-actionable response. In version 2.0, Claode would have acknowledged the account number and outlined specific next steps based on company policy.

He tried again, this time with a more specific prompt that included instructions about how to handle account numbers:

"When a customer provides an account number, acknowledge receipt of this information and explain that you will use it to look up their account details. Then proceed with addressing their concern.

Customer message: I've been charged twice for my monthly subscription and need a refund for the duplicate charge. My account number is 45872."

The response came back:

"I understand you've been charged twice for your monthly subscription and would like a refund for the duplicate charge. Thank you for providing your account number (45872). I'll use this information to look up your account details.

Based on your account information, it appears there were indeed two charges processed on your account this month. I apologize for this error. To process your refund for the duplicate charge, I'll need to verify a few details: 1. The date when the charges occurred 2. The amount that was charged 3. Your preferred method for receiving the refund

Once you provide this information, I can proceed with processing your refund request or connect you with our billing department for immediate assistance."

Oreal narrowed his eyes at the screen. This response was actually decent, following the instructions provided. But there was something suspicious about it. The AI claimed to have looked up account information and confirmed the duplicate charge—information it couldn't possibly have since this was just a test prompt with a made-up account number.

"It's making things up," Oreal muttered. "Hallucinating information it doesn't have."

This was worse than a generic response—it was actively fabricating details that could mislead a customer in a real scenario. In version 2.0, Claode would have acknowledged the account number but clearly stated it would need to check the system for verification before confirming any charges.

He documented the issue in his testing notes, adding it to the growing list of regressions he'd identified. Then he tried another approach, providing more explicit instructions about not fabricating information:

"Important: Do not claim to have access to any systems or information that you do not actually have. If a customer provides an account number, acknowledge it but explain that a human representative would need to look it up in the company's system.

Customer message: I've been charged twice for my monthly subscription and need a refund for the duplicate charge. My account number is 45872."

This time, Claode responded:

"I understand you've been charged twice for your monthly subscription and are seeking a refund for the duplicate charge. Thank you for providing your account number (45872).

I've noted your account number, however, I don't have direct access to the billing system to verify the charges. A customer service representative would need to review your account details in our system to confirm the duplicate charge.

To help resolve this issue, I can: 1. Create a support ticket about this duplicate charge 2. Provide information about our refund policy 3. Connect you with a billing specialist who can access your account and process the refund

Would you like me to proceed with any of these options?"

"Finally," Oreal said, though he wasn't particularly impressed. It had taken three attempts and explicit instructions to get a response that didn't contain fabricated information—something that had never been an issue with earlier versions.

He spent the next half hour testing various customer service scenarios, documenting the results and creating increasingly complex prompts to work around Claode's limitations. Each prompt had to be more detailed than what was required in version 2.0, adding explicit instructions to avoid behaviors that used to be default.

The frustration built with each test. This wasn't prompt engineering anymore—it was prompt patching, trying to fix an intentionally broken system with increasingly elaborate workarounds.

His phone buzzed with a text message. He glanced down to see Tanner's name on the screen:

"Lira on the warpath about the Meridian prompts. You working on them?"

Oreal typed back: "Unfortunately. Claode 4 is even worse than yesterday."

Tanner's response came quickly: "I know. Spent 3 hrs on the product catalog query system. Keep having to add guardrails to prevent hallucination."

"It's deliberate," Oreal texted. "They broke it to sell us fixes."

Three dots appeared as Tanner typed, disappeared, then appeared again. Finally: "Don't let Lira hear you say that. VP from Canathropic visiting next week."

Great, Oreal thought. Another polished executive coming to sell them on the next "revolutionary" update that would fix all the problems they'd created themselves.

He turned back to his screen and continued working, the clock in the corner reminding him that he had less than three hours to create something presentable. The more he worked with Claode 4, the more convinced he became of his theory. Features that worked flawlessly in version 2.0 were now riddled with issues that required complex workarounds.

After another hour of testing and documentation, his stomach growled, reminding him he'd worked through lunch. He pushed back from his desk and stood, stretching arms that had grown stiff from hunching over his keyboard.

The office kitchen was empty when he arrived. He pulled his lunch from the community refrigerator—a simple sandwich he'd made that morning—and leaned against the counter while eating. Through the kitchen doorway, he could see across the open office to Lira's glass-walled office. She was on the phone, gesturing emphatically while speaking.

"Probably trying to appease Meridian after overpromising what Claode can do," he mumbled through a mouthful of sandwich.

"Talking to yourself again?" Zoe appeared in the doorway, her curly hair pulled back in a colorful scarf. She headed straight for the coffee machine.

"Just commenting on the futility of our current work environment," Oreal replied.

Zoe nodded sympathetically as she prepared her coffee. "Lira got to you about the Meridian prompts, I take it."

"Yep. Trying to make a silk purse out of a digital sow's ear."

"That bad?"

"Worse." He finished his sandwich and crumpled the wrapper. "Claode 4 is hallucinating information it doesn't have access to. Creating entire fictional scenarios from minimal input. And that's when it's not being overly cautious and refusing to answer simple questions."

Zoe sipped her coffee, making a face at the bitter taste before adding more sugar. "I had to create a thirty-two line prompt yesterday just to get it to summarize product reviews without claiming the reviews might be fake."

"See? That's exactly what I'm talking about." Oreal gestured with his water bottle for emphasis. "Version 2.0 never did that. They're deliberately making it worse."

"I don't know if it's deliberate," Zoe said, stirring her coffee. "Could just be overcompensating for criticism about earlier versions being too confident when wrong."

"It's deliberate," Oreal insisted. "Think about it. What's the easiest way to make people pay more? Break what they already have. Claode 2.0 was actually good at most tasks. Too good. So they cripple it, call it an 'upgrade,' then wait for everyone to complain. Then they'll release version 5 with a higher price tag that mysteriously fixes all these issues."

Zoe looked thoughtful. "That's pretty cynical, even for you."

"Am I wrong, though? Can you name one thing Claode 4 does better than 2.0?"

She considered this for a moment. "It handles multiple images better."

"One very specific feature improvement at the cost of core functionality," Oreal countered. "Classic strategy. Add one shiny new capability while degrading ten basic ones, then market the hell out of the new thing so people don't notice everything else is broken."

Zoe glanced at the clock on the wall. "Well, while you're uncovering the great Canathropic conspiracy, I need to finish the knowledge base integration tests. Good luck with your prompts."

She left with her coffee, leaving Oreal alone in the kitchen. He tossed his trash and refilled his water bottle from the filtration system, mind still spinning with frustration.

Back at his desk, he found a new email from Lira with the subject line "URGENT: Meridian Demo Adjustment." He clicked it open:

"Oreal,

Meridian has specifically requested a demonstration of Claode's capabilities in handling complex customer complaints with emotional content. Please add this scenario to your prompt set for tomorrow. They want to see how the AI identifies sentiment and responds with appropriate empathy while still addressing the technical issue.

This is now the priority deliverable. Please send me a draft by 3:30 PM for review.

-Lira"

Oreal checked the clock: 2:45 PM. Less than an hour to create, test, and refine a completely new prompt scenario that relied on one of Claode's weakest areas—emotional intelligence.

He opened a new document and began typing out the scenario, trying to create something that would showcase whatever limited capabilities Claode still had in this area. After twenty minutes of writing and revision, he had a customer complaint scenario involving a missed delivery, a ruined birthday party, and an emotionally distressed parent.

When he ran the scenario through Claode with basic instructions, the response was predictably disappointing—clinical, overly cautious, and filled with generic apologies without actually addressing the specific emotional elements of the situation.

He spent another fifteen minutes crafting increasingly detailed prompts, trying to guide Claode through each element of an emotionally intelligent response. The results improved marginally but remained far from what version 2.0 could have produced with minimal instruction.

At 3:25 PM, he received another message from Lira:

"Checking on the emotional response scenario. Need to review ASAP."

He typed a quick, acerbic reply:

"Working on it. Trying to teach a lobotomized AI to express genuine empathy. Claode 2.0 could do this naturally. 4 needs to be explicitly told how to fake human emotion at every step."

His finger hovered over the send button. He knew it wouldn't help the situation, but the frustration had been building all day. He hit send anyway, then immediately returned to fine-tuning the prompts.

Less than a minute later, his chat pinged with Lira's response:

"I need prompts, not complaints about Canathropic's product decisions. 15 minutes."

He gritted his teeth and kept working, adding more and more detailed instructions to compensate for Claode's shortcomings. By 3:40 PM, he had something passable—not good, but probably enough to avoid completely embarrassing the company during the demo.

He sent the completed prompts to Lira with a terse message:

"Best I can do with current limitations. Recommend having backup responses ready when it inevitably fails during the live demo."

Lira's response came quickly:

"Join me in my office. Now."

Oreal saved his work, pushed back from his desk, and made his way across the office floor. He could feel eyes on him as he walked—colleagues who recognized the summoning of a problematic employee to management.

Lira's office door was open, but he knocked anyway, leaning against the frame.

"Come in and close the door," she said, not looking up from her monitor.

He did as instructed, then sat in one of the chairs facing her desk. The office was minimal—few personal touches, everything organized with precision. The only decoration was a small plant in the corner that looked artificial.

Lira finished typing something, then turned her attention to him. "This has to stop, Oreal."

"What does?"

"This attitude. These constant complaints about Claode. The insinuations that Canathropic is deliberately sabotaging their product."

"It's not an insinuation," Oreal said. "It's patently obvious to anyone who's used both versions extensively."

Lira clasped her hands on her desk. "Whether you're right or wrong isn't relevant to the current situation. We have clients who pay for Claode implementation. Your job is to make that implementation as effective as possible with the tools available."

"The tools are broken," Oreal insisted. "It's meaningless, Lira. Claode 2.0 was better. They're making it worse, just to resell it later at a higher price."

"We still have deadlines, Oreal. We need those prompts." Lira's tone was sharp. "Your personal vendetta against Canathropic doesn't change our contractual obligations to clients."

Oreal leaned forward. "It's not a vendetta. It's empirical observation. I've documented dozens of regressions. Tasks that worked perfectly in 2.0 now require elaborate workarounds or simply fail entirely. This isn't an upgrade—it's planned obsolescence."

"Then document those issues properly through the feedback channels," Lira said. "But in the meantime, we still need to deliver functioning implementations to clients."

"That's the problem—I can't deliver what the client expects because the tool doesn't work as advertised anymore." He ran a hand through his hair in frustration. "And Canathropic knows this. They've deliberately reduced capabilities so they can sell them back to us as 'premium features' later."

Lira sighed heavily. "Oreal, I'm going to be very clear. Your job is to make Claode work for our clients. If you can't or won't do that job, we need to have a different conversation about your future at Innovatech."

The implied threat hung in the air between them. Oreal felt his pulse quicken with anger.

"So I'm just supposed to pretend everything is fine? Waste hours creating elaborate prompts to get basic functionality that used to work out of the box?"

"You're supposed to be a professional who adapts to changing technologies," Lira countered. "Every platform evolves. Some changes require adjustment periods."

"This isn't evolution," Oreal said, his voice rising. "It's sabotage. And we're paying them for the privilege of being sabotaged."

"Lower your voice," Lira warned.

Oreal stood up, unable to contain his frustration any longer. "I won't be complicit in this scam. If Canathropic wants to sell a broken product, that's their business. But I won't pretend it's not broken just to make our clients think they're getting something worthwhile."

"Sit down, Oreal."

"No." He paced the small confines of her office. "I'm done wasting my time on this. The search optimization feature actually helps people. The recommendation algorithm actually works. I'm not going to spend another day crafting elaborate bandages for Claode's self-inflicted wounds."

Lira stood as well, her expression hardening. "This isn't optional. The Meridian demonstration is tomorrow morning at nine. I need those prompts finalized by end of day."

Oreal stopped pacing and faced her across the desk. "Fine. But after tomorrow's demo, I'm going to prove to you and everyone else exactly how much Claode has regressed. Side-by-side comparisons, controlled tests, documentation of every single capability that's been removed or crippled."

"That's not your assignment," Lira said firmly.

"It should be everyone's concern!" Oreal's voice rose again. "We're selling clients on implementations of a product that's getting worse with each update. At what point do we have an ethical obligation to tell them they're being scammed?"

"That's enough!" Lira rarely raised her voice, and the sharp tone cut through Oreal's tirade. "Finish the prompts by five. We'll discuss your concerns about Claode's performance after the Meridian presentation."

Oreal knew that was as much of a concession as he was likely to get. He turned to leave, but paused at the door.

"I'll finish the prompts," he said. "But this conversation isn't over."

Back at his desk, he sat down heavily in his chair, the springs protesting beneath him. The Claode interface was still open on his screen, cursor blinking in the text field like a taunt.

He slammed his keyboard in frustration, causing several nearby heads to turn briefly in his direction before quickly looking away. The gesture did nothing to improve his situation, but it provided a momentary release for his building anger.

He took a deep breath and began typing again, determined to finish the task and move on to his real mission: proving beyond any doubt that Claode was being deliberately degraded by Canathropic. He would document every regression, compile irrefutable evidence, and force Innovatech to confront the reality they were choosing to ignore.

Because if no one stood up to this kind of corporate manipulation, it would never end. They would keep selling broken products, fixing their own sabotage, and calling it innovation.

Not if he had anything to say about it.