Claude Opus 5 Vibe Coding: Gamer Builds AI Strategy App for The Finals

A new PCMag report tells an unusual story. A player of Embark Studios' chaotic free-to-play shooter The Finals used Claude's latest model, Opus 5, to vibe-code a full study app called FISH   short for Finals Interactive Study Helper. The app never touches the game's files and doesn't run alongside it. Instead, it generates random match states and quizzes the player on the correct decision.

The report matters because this is not another "I built a to-do list with AI" anecdote. It is a detailed, hands-on look at where AI coding stands in mid-2026 — specifically Opus 5's agent capabilities, its real-world limitations, and how the gap with OpenAI's GPT-5.6 is narrowing.

Below, we break down what FISH does, what building it with Opus 5 was actually like, and what the project reveals about the current state of vibe-coded software.

What FISH Actually Does

The Finals is a team-based first-person shooter where three-player squads fight over cashouts across destructible arenas. It is more chaotic and dynamic than Battlefield or Call of Duty, and success depends heavily on split-second, in-the-moment decisions.

FISH is built to sharpen those decisions. It has two core functions.

1. Match Simulation

FISH generates random but realistic game states across different rounds and phases. Each simulation includes:

• Gadgets and weapons equipped by every player
• Player states — alive, dead, and respawn timer
• Live scoreboard and team standings
• Number of respawn coins available
• Team composition (light, medium, heavy builds)
• Team-wipe costs
• A pannable, zoomable map plotting objectives, players, and statues

FISH then delivers an "answer"   what the correct play should be in the state it generated.

2. Memorization Drills

The app also trains recall. It shows three random images of in-game items — gadgets, interactive objects, specializations, maps, and map variants  without names, descriptions, or stats. Players quiz themselves, then reveal the information to see what they actually knew. For maps, FISH helps reinforce layouts and which modifiers each map can roll.

Why this matters: The memorization side is straightforward flashcard logic. The simulation side is far more ambitious   and it is where FISH tries to do something genuinely difficult.

Two Types of Answer

FISH separates its advice into two tiers.

Rulebook answers are grounded in the game's basic rules and mechanics. For example: if it is late in Round 1, your team is in third place, and only one cashout is active, the rulebook says to steal that cashout.

Pro strategy answers are more specific, more nuanced, and far harder to generate. This is where the project gets interesting.

A Concrete Example: How FISH Thinks

Consider a randomly generated state. It is the final round. Your team is ahead. A nearby cashout is controlled by the enemy team   and in the simulation, the enemy is down one player.

Rulebook answer: Push the cashout.

Pro strategy answer:

• Take the fight, but now is the moment for your light teammate to engage aggressively, because the enemy is a player short.
• That light teammate should keep their glitch grenade ready to counter the enemy's equipped gateway gadget.

This kind of context-specific advice comes from FISH's library of roughly 70 strategy principles, each applicable to different situations. The rulebook and pro strategy frequently disagree   and that disagreement is part of the value.

Important distinction: FISH is not a cheat. It does not interact with game files, does not read live state, and does not run concurrently with the game. It is a training and study tool   closer to a chess strategy workbook than an aimbot.

Building With Opus 5: The Great, the Good, and the Limit-Draining

FISH was built using Opus 5 on Anthropic's Claude Max 5x plan, which costs $100 per month. The reporter had previously used Fable 5 (the prior Opus generation), and the contrast is instructive.

What Worked Well

1. Efficiency gains. Compared with Fable, Opus 5 allows substantially more prompts before hitting the five-hour usage limit. For coding projects, that is a meaningful step forward.

2. Complex tasks were surprisingly easy.

• A model to simulate The Finals matches took only a few prompts.
• A pannable, zoomable map for displaying simulated states was similarly straightforward.
• Most impressive of all: building a system that gives informed advice on randomly generated game states was "surprisingly easy," according to the report.

3. YouTube transcript analysis. This is the technical heart of the project, and we cover it in the next section.

What Was Frustrating

1. The five-hour limit burns fast with agents. By default, Claude may spin up 10 or 20 agents for a single task, exhausting the five-hour window in minutes. During an audit of FISH, the reporter stepped away for a few minutes and returned to find the limit nearly gone   worsened by the fact that agent failures often require redoing work.

The reporter contrasts this with ChatGPT's default maximum of three agents, which they generally prefer.

2. Inconsistent agent behavior. Sometimes Opus 5 launches only one or two agents that run for 20 or 30 minutes, then another one or two   working through the entire five-hour window without ever exhausting the limit. The reporter observed this while debugging FISH's parsers. Other times, it floods the task with agents and drains the limit instantly. The inconsistency makes planning difficult.

3. Web search weaknesses. FISH needed large amounts of information about The Finals' items and mechanics, all readily available on the official site and the community wiki. Opus 5's agents routinely failed to extract information from these easily accessible pages, forcing the reporter to provide links and quoted text manually. GPT-5.6's agents did not have the same problem.

4. Memory instructions were frequently ignored. The reporter set up several memories, including:

• A directive to be concise   Claude tends toward verbosity and meta-commentary   yet Opus 5 continued to bloat files.
• A rule to stop at 80% context window usage and update a handoff document for a fresh session. Opus 5 kept stopping at random points, unable to accurately gauge how full its context actually was.

Practical takeaway: Opus 5 is powerful, but agent management and limit discipline still require active human oversight. Anyone starting a large project should explicitly cap agent count in memory settings and expect to supervise long runs.

The Technology That Sets FISH Apart: Turning YouTube Guides Into a Strategy Bible

The most compelling part of the project is how FISH generates pro-level advice. The process worked like this:

1. Transcript extraction. The reporter pulled transcripts from hundreds of YouTube videos by experienced The Finals players.
2. Analysis. Opus 5 was asked to analyze every transcript and produce a "strategy bible."
3. Processing. Dozens of agents and millions of tokens later, the bible was complete.
4. Principle distillation. Opus 5 distilled the bible   along with its citations   into discrete principles that it could tag against relevant game states and serve as advice.

The result is a library of roughly 70 principles, each applying to different in-game situations. Advice is therefore neither repetitive nor obvious from the rules alone.

The reporter emphasizes that the entire pipeline was remarkably simple to set up, even though the processing itself took Opus 5 considerable time.

Industry context: This is a notable milestone. A few years ago, extracting structured strategic knowledge from hundreds of hours of video would have required a development or data team working for over a week. Today, an individual user can do it in an evening with a frontier model   and then wrap that knowledge into an interactive application. The bottleneck has shifted from building the technology to curating the source material.

Does FISH Actually Improve Gameplay?

According to the report, the act of building the app was itself instructional. The reporter cites several concrete improvements:

• Faster recall. Gadgets, interactive items, maps, map modifiers, and specializations now come to mind more readily.
• Quicker analysis. Reading a situation and responding to it has become faster.
• Team-wipe timing. FISH repeatedly reinforces that wiping and respawning is sometimes faster than trying to revive teammates.
• Goo placement. Placing goo one level below a cashbox can block an enemy RPG blast from knocking the cashout out of reach.
• Steal vs. wipe. The reporter now better understands when to play for a cashout steal versus wiping the enemy team first.

That said, realism matters. With only a few dozen hours in The Finals, the reporter still describes themselves as "a total novice." One week with FISH has produced clear learning and genuine enjoyment, but major improvement will still require many more hours and matches.

What this means for readers: FISH demonstrates that AI-generated training tools can improve real-world skill  but they are not a magic wand. They cannot replace deliberate practice and match time, but they can structure and accelerate the learning process.

Opus 5 vs. GPT-5.6: The Gap Is Closing

This report carries particular weight because the reporter has previously compared GPT-5.5 against Opus 4.6, and Fable against GPT-5.6. In every prior comparison, OpenAI's models came out clearly ahead.

Opus 5 changes that picture   though not completely

CLAUDE OPUS 5 VS. GPT-5.6 PERFORMANCE & BEHAVIOR COMPARISON
AREA	OPUS 5	GPT-5.6
Efficiency (tokens/session)	Significantly more efficient than Fable	—
Agent Management	Default 10–20 agents; drains limits quickly	Default max 3 agents; more controlled
Web Search	Failed on simple, accessible pages	No comparable issue reported
Memory Adherence	Frequently ignored instructions	—
Coding Quality	Strong, especially on complex systems	Still delivered the best results on the reporter's projects
Overall Intelligence	Competitive with Fable 5	—

The reporter's conclusion is measured. GPT-5.6 still tends to produce the best results for their coding projects, including FISH. But every project is different, and not every task is a natural fit for OpenAI's models. With its efficiency gains and Fable 5-level intelligence, Opus 5 is "genuinely worth consideration"   and at minimum, worth trying.

Five Lessons for Vibe Coders

The FISH project offers several practical takeaways for anyone building with frontier AI models.

1. AI can now do knowledge synthesis, not just code. Analyzing hundreds of video transcripts into structured, retrievable strategy is knowledge engineering, and a frontier model handled it with minimal scaffolding.

2. Agent limits need active management. Opus 5's agents are powerful but can burn through usage windows in minutes. Cap agent counts explicitly and monitor long runs.

3. Memory is not reliable on its own. Setting an instruction in memory does not guarantee adherence. Restate critical instructions in each prompt for high-stakes work.

4. Web research still needs human oversight. Even when source pages are public and well-structured, AI agents can fail to extract the right information. Plan to verify and supply links manually.

5. Vibe coding does not mean zero effort. The reporter audited output, supplied links and quotes, debugged parsers, and monitored agent behavior throughout. AI amplifies effort   it does not replace it.

What Comes Next

FISH remains an active project. Its 70 principles will likely grow, and the reporter plans to log more hours in The Finals to measure the app's real-world impact over time.

The bigger picture is what matters for the industry. The narrowing gap between Opus 5 and GPT-5.6 is good news that extends beyond Anthropic   stronger competition means more capable tools, lower costs, and better outcomes for the people building with them.

For readers considering their own AI-assisted projects, the practical advice is simple: do not default to a single model. If you are starting a coding project, give Opus 5 a genuine try. It may turn out to be the better fit for your specific task.

The Bottom Line

FISH proves that vibe coding has moved past simple websites and to-do apps. A motivated individual can now build complex simulations, knowledge-based strategy systems, and interactive training tools   provided they have the right model, sufficient patience, and a willingness to stay involved.

Opus 5 is not perfect. Its agent management is uneven, its web search is inconsistent, and its memory adherence needs work. But it has shown that Anthropic is closing the distance with OpenAI   and that is a competition from which everyone building with AI stands to benefit.

Source: PCMag   "I Vibe Coded an App With Opus 5 to Get Better at The Finals, and It's Working"