Timing is a funny thing. The day we launched SEOgent I saw this post on X from Uncle Bob Martin:
"Don't burn tokens on fixing things that the AI isn't good at fixing. Instead have the AI build a tool that can fix them quickly without burning tokens."
That exactly encapsulated what I am hoping to accomplish with SEOgent. My goal from the initial planning stages was to leverage existing tools and methodologies for website crawling and analysis to help AI leverage that data to make better decisions on what and how to fix SEO issues on websites.
I decided to take a deep dive into this line of thinking and see what other experts in the development and AI worlds have to say. Was I on the right track? Do tools that do the heavy lifting behind the scenes help or hinder your local AI workflow? Here's what I found.
Uncle Bob's tweet didn't come out of nowhere. A few weeks earlier he shared a longer reflection on his experience using AI for coding. He compared himself to a carpenter who's been using hand tools his whole career and suddenly gets access to a shop full of power tools. The power is undeniable, he said, but so are the risks and time sinks. He wasn't even convinced his project would be further along than if he'd just written it all by hand.
I think most of us who have been building with AI have had that moment. You're deep into a conversation with Claude or ChatGPT, you've burned through a pile of tokens, and you stop and think "I could have just written this myself in half the time." The answer isn't to stop using AI. The answer is to use it smarter.
Developers are burning through tokens at an alarming rate, often on tasks that don't need AI at all. One developer found that a single prompt asking Claude to commit and push code consumed 25% of his entire session quota. He started doing Git operations manually just to save tokens for actual coding.
Another developer recently posted on DEV Community how an AI coding agent would read 30-40 files at the start of every session just to orient itself. The same files. The same discoveries. The same wasted tokens. Every single time. By addressing that repetition the developer cut input tokens per query from around 8,200 down to 2,100 — a 65% reduction without changing models.
The pattern is clear: most token waste comes from repetition and from throwing AI at problems where a simple script or existing tool would be faster, cheaper, and more reliable.
I wanted to see if this "build a tool instead of burning tokens" idea was just Uncle Bob being Uncle Bob (IYKYK), or if there was a broader trend here. Turns out it's everywhere.
Addy Osmani, an engineering leader at Google, published a detailed breakdown of his LLM coding workflow heading into 2026. One section jumped out at me. He highlights Claude Skills as a game-changer specifically because they turn "fragile repeated prompting into something durable and reusable." Instead of telling the AI the same thing every session, you capture that knowledge once and let it apply automatically.
He also advocates for being surgical about what you feed the AI. Explicitly tell it what not to focus on to save tokens. Only include the code portions relevant to the task. Don't dump your entire codebase into the context window and hope for the best.
This is the same philosophy behind SEOgent. Rather than asking an AI to crawl a website and figure out what's wrong from scratch every time — burning massive tokens reading HTML, following links, interpreting page structure — SEOgent does the crawling and analysis with a purpose-built tool and then hands the AI clean, structured data. The AI spends its tokens on what it's actually good at: analysis and recommendations.
Anthropic's engineering team published some findings that really drove this home for me. Their research into code execution with MCP (Model Context Protocol) showed that when agents load all tool definitions upfront, a simple request like "list the 10 most recent issues from my GitHub repo" consumed 102,000 tokens. Not because the task was complex, but because the model received metadata for 114 tools — most of which had nothing to do with the request. Even saying "hello" burned over 46,000 tokens.
Their solution? Instead of loading everything into context, let agents discover and call only the tools they need. Token usage dropped from 150,000 to 2,000. A 98.7% reduction.
The really interesting part is their concept of agents persisting code as reusable functions. Once an agent develops working code for a task, it saves that implementation as a "skill" for future use. Over time the agent builds a toolbox of capabilities. It's essentially Uncle Bob's tweet turned into an engineering pattern: have the AI build the tool once, then use the tool going forward instead of burning tokens to figure it out again.
That's exactly why we built (and show you how) to create an SEOgent skill to utilize our CLI tool. It reduces all the noise and waste and gets right to analyzing and solving the issues on your website.
Birgitta Böckeler at Thoughtworks ran experiments pushing AI autonomy in code generation and landed on a great insight: "Think about how you can maximize the abstraction level of the code you are generating with AI, to take advantage of the speed and reliability of deterministic software as much as possible."
She specifically suggests generating a script or codemod instead of letting AI do the full work itself. For bootstrapping their applications they used a shell script rather than having the LLM do it — because there's already a CLI that creates an up-to-date, properly structured application. Why burn tokens on something a deterministic tool already handles perfectly?
Our thoughts exactly. Why waste tokens crawling a website and running tests that AI doesn't need to run. Free up your tokens to be better spent on the tasks where they can return the most benefit.
Their experiments also found that providing the AI with a reference application via MCP was far more effective than explaining everything in natural language. The patterns stayed consistent, the code compiled, and the results were repeatable. Give AI structured inputs and it produces structured outputs. Give it vague prompts and you get vague results.
This is exactly the gap SEOgent was built to fill.
The typical approach to AI-powered SEO analysis goes something like this: feed a URL to an AI, ask it to crawl the site, have it analyze the HTML, then ask it what's wrong. Every step burns tokens. The AI is reading raw HTML, following links, interpreting page structure, deduplicating content — all tasks that traditional crawlers have been doing reliably for decades. You're wasting a wallet full of tokens for work that a Go binary can handle in a fraction of the time at a fraction of the cost. On top of that, SEOgent runs industry-standard SEO tests and analysis with every scan — checking things like meta tags, heading structure, broken links, redirect chains, schema markup, and Core Web Vitals. You don't need to spend time training your AI to understand SEO best practices or figuring out the right prompts to get a proper audit. That part's already handled before your AI even sees the data.
SEOgent flips that. The crawler does what crawlers do best — systematically visiting pages, extracting metadata, checking response codes, mapping internal links, analyzing page structure. It produces clean, structured, machine-readable data. Then your AI agent — whether it's Claude, GPT, or whatever you're working with — receives that structured data and spends its tokens on the high-value work: identifying patterns, prioritizing issues by impact, generating specific fix recommendations, and explaining the "why" behind each suggestion.
Your AI agent gets busy on the real issues, the problems plaguing your website, instead of wasting time and tokens trying to find the problems in the first place.
The developers getting the best results from AI right now aren't the ones throwing more tokens at problems. They're the ones building better workflows — using purpose-built tools for the tasks those tools are good at, encoding repeated knowledge into reusable skills and scripts, and reserving AI for the work that actually requires intelligence and judgment.
Uncle Bob nailed it in one sentence. The rest of the industry is arriving at the same conclusion through hard-won experience and expensive API bills.
The smartest way to use AI is to first build the tools that make AI smarter. That's what SEOgent is. Not an AI that does SEO — a tool that makes your AI better at SEO.