Codify Before You Build: How to Actually Use AI in Corporate Engineering

If you spend any time on LinkedIn or engineering forums these days, you are bombarded with the same message: "AI is going to change everything." We see demos of generative design algorithms creating impossible geometries and language models writing Python scripts in seconds. But when you sit down at your desk on a Tuesday morning, staring at a backlog of ECOs (Engineering Change Orders), a stack of supplier deviation requests, and a looming design review, the "AI revolution" feels very far away.

The disconnect between the hype and our daily reality is palpable. Most mechanical engineers in corporate environments aren't trying to invent a new method of propulsion; they are trying to get a drawing package approved before the Friday deadline.

Recently, a fascinating video surfaced detailing how non-technical employees inside Anthropic—the $380 billion company behind the Claude AI models—actually use their own tools [1]. The insights from this video are profound, not because they show off futuristic technology, but because they reveal a highly practical, systematic approach to work that translates perfectly to the mechanical engineering discipline.

It turns out, the secret to leveraging AI isn't about learning the underlying theories of neural networks. It is about identifying, describing, and solving your own critical workflows.

The Myth of the "Do Everything" AI

The biggest mistake engineers make when approaching AI is treating it like a junior engineer who can just "figure it out." We throw a massive, ambiguous problem at a language model—"Optimize this assembly for manufacturing"—and are disappointed when it returns generic, unhelpful advice.

The reality inside companies like Anthropic is very different. They do not use AI as a monolithic problem solver. Instead, they build "micro-tools."

A micro-tool is a highly focused, single-purpose system designed to eliminate one specific bottleneck in a workflow. It doesn't try to do your whole job; it tries to do the 80% of a specific task that is routine, repetitive, and doesn't require your highest level of engineering judgment.

While the video specifically highlighted workflows using Claude Code, it is crucial to understand that the landscape of AI tools is vast and rapidly evolving. Whether you are using Claude, OpenAI's ChatGPT, Microsoft Copilot, or specialized engineering AI platforms, the underlying principles remain exactly the same. The tool is just the engine; the value comes from how you build the track.

Case Study 1: The "Department of No" Becomes a Thought Partner

One of the most compelling examples from the Anthropic video involved their legal department. In most corporate environments, legal (much like Quality Assurance or Document Control in engineering) is often viewed as the "department of no." They are the bottleneck where fast-moving projects go to wait for manual review.

The Associate General Counsel at Anthropic faced a massive backlog of marketing materials requiring legal approval. Instead of hiring more lawyers or working longer hours, he built a self-service micro-tool pinned directly in the company's Slack workspace.

Here is how the workflow operates:

A marketer pastes their proposed content into the tool.
The AI checks the text against the company's codified legal guidelines.
It flags specific issues (e.g., missing trademark symbols or risky claims).
It assigns a risk level and tells the marketer exactly how to fix the issues.
Once the marketer makes the corrections, the tool generates a clean, formatted summary for the legal team to do a final, rapid review.

The Engineering Translation: Supplier Deviation Requests

How does this apply to mechanical engineering? Consider the process of reviewing Supplier Deviation Requests (SDRs).

When a supplier cannot meet a specific tolerance on a drawing, they submit an SDR. An engineer must then review the request, check the assembly stack-up, evaluate the functional impact, and approve or reject it. It is tedious, time-consuming work.

Applying the micro-tool philosophy, you could build a system where the supplier submits the SDR data into a structured form. The AI tool, armed with your codified rules (e.g., "If the deviation is on a non-mating surface and is less than 0.1mm, flag as Low Risk"), pre-screens the request.

The tool handles the 80% of the routine checks. It flags the critical dimensions that actually require your engineering judgment—the 20%—and presents them to you in a clean, standardized format. You are no longer digging through PDFs; you are making high-level engineering decisions.

Case Study 2: Scaling Output Without Scaling Headcount

The second example focused on a growth marketer who was operating as a one-person team during a period of massive company expansion. He needed to generate hundreds of variations of ad copy and visual creatives, a task that would normally require a team of copywriters and designers.

He solved this by building two specific systems. First, he used AI to write a custom script for his design software (Figma) that automatically generated visual permutations of an ad with a single click, reducing a 30-minute task to 30 seconds. Second, he created a custom AI skill that cross-referenced campaign data against brand guidelines and platform best practices to instantly generate optimized ad copy.

He didn't just ask the AI to "write an ad." He built a repeatable system that codified his marketing expertise.

The Engineering Translation: Automated Test Reporting

Think about the hours spent compiling test reports. You run a thermal cycle test on a prototype, extract the data from the thermocouples, paste it into Excel, create the plots, and then write a 20-page Word document explaining the results.

Instead of doing this manually, you can build a micro-tool. You write a Python script (which an AI can help you write, even if you aren't a programmer) that automatically ingests the raw CSV data from the test equipment and generates the plots. You then pass those plots and the key metrics to an AI model, along with a strict template of how your company formats test reports.

The AI generates the boilerplate text, inserts the plots, and flags any anomalies that exceed your predefined test limits. You review the final document, add your specific engineering conclusions, and sign off. You have just turned a two-day task into a two-hour task.

The Core Pattern: Codify Before You Build

The underlying pattern in both of these examples is the most important lesson for corporate engineers: Codify your expertise first.

AI cannot replace your engineering judgment, but it can scale it. However, to scale your judgment, you must first write down your rules. The AI needs to know your boundaries, your formatting requirements, and your decision-making criteria.

If you want an AI to help you review drawings, you first need a written, unambiguous checklist of what constitutes a "good" drawing in your organization. If you want it to help you select materials, you need to codify your constraints (e.g., "Must be RoHS compliant, yield strength > 250 MPa, cost < $5/kg").

You are not replacing the expert; you are building internal capacity to handle the repeatable, boring work so the expert can focus purely on tasks that require actual human judgment.

Your Action Plan: Start Small

The transition to AI-augmented engineering doesn't happen with a massive, top-down corporate initiative. It happens when individual engineers start solving their own micro-problems.

Here is your action plan for this week:

Pick ONE Workflow: Choose one specific task you do every week that drains your energy. Do not pick "designing a new product." Pick "formatting the weekly project status update" or "checking BOMs for obsolete part numbers."
Write Down the Steps: Document exactly how that task is completed from start to finish. What data do you need? Where do you get it? What rules do you apply?
Identify the 80%: Figure out which parts of the task are routine, repetitive, and don't require your best engineering thinking. This is the data gathering, the formatting, and the basic rule-checking.
Identify the 20%: Figure out which parts of the task carry a high "risk of error" and absolutely require your human judgment and expertise. This is the final approval, the complex stack-up analysis, or the safety-critical decision.
Build the System: Use an AI tool (whether it's Copilot, ChatGPT, or a custom script) to build a micro-tool that handles the 80%. Keep yourself firmly in the loop to review and approve the final 20%.

Stop waiting for the perfect, all-knowing engineering AI to arrive. The tools you need to eliminate the worst parts of your job are already here. You just need to codify your rules and build the system. Take action today, pick one small workflow, and reclaim your time.

References

[1] Marchese, A. (2026, March 28). How Anthropic Employees ACTUALLY Use Claude Code. YouTube. https://www.youtube.com/watch?v=pRfIWmddRsE