AI for R&D

9 minute read•Last updated November 12, 2026

Recorded live · Software R&D team

How engineers get expert results from AI: context, Claude Code, skills and subagents

I taught this class live to a team of software R&D engineers. It covers how a model reads your words, why context decides every answer, why engineers should work in Claude Code, and how skills, subagents and index files teach the AI your team's way of working. The written version is below, so you can skim it or come back to it later.

1. A model is a brain in a jar

A large language model is like a brain in a jar. It takes in one thing, tokens, and it gives back one thing, tokens. Think of a token as a word. It is not exact, but it is a good mental model.

Claude.ai is an app built around that brain. You send a prompt, which is an instruction in plain text. The model answers. Your prompts and its answers build up the chat. Everything in the chat is the model's context: all the words it can see right now.

First tip: write to AI in English. English uses fewer tokens than most other languages. Other languages work too, but they cost more tokens for the same meaning. On long engineering tasks, English is faster, cheaper, and a little better.

2. Context decides the answer

The model answers only from its context. Give two people the same model and the same context, and they get the same quality of answer. Manage the context and you control the output.

At the time of filming, the top Claude models hold about one million tokens of context. Past that, the oldest words fall out, first in, first out, and the model never sees them again.

The real limit is lower. Around 300,000 to 400,000 tokens, the AI starts to get distracted. An instruction it would follow perfectly in a short chat now has a smaller chance of being done right. So the rule is simple: when a chat passes 300,000 tokens, start a new one. Bring a summary or the last answer with you.

Claude.ai compacts long chats on its own. It summarizes the old chat and keeps going from the summary. Every round loses detail. When the topic changes, start a new chat instead.

3. Tools turn a model into an agent

Give the brain tools and it becomes an AI agent. Tools are its body. Web search lets it pull fresh results into its context. In Claude Code, it can read, write and search the files on your computer and run commands. With only a few tools (bash, read, write, glob and grep) and the right permissions, an agent can do a huge amount of real work.

An agent is a model plus a harness. The model is the brain you pick. The harness is the software around it: its instructions, tools and rules. Claude Code is a harness.

4. Pick the strongest model for hard work

At the time of filming, Fable is the strongest Claude model and Opus 5.5 comes next. Use Fable for hard work until you hit your usage limit. Skip it for simple questions. You do not need it there, and you will wait longer. For coding, pick the top model and the max effort setting. It uses up your limits faster, but the code comes out right and you spend less time fixing it.

Models change fast. To see which one leads today, check the public leaderboard at arena.ai.

5. System instructions get the most attention

Every major AI tool lets you set system instructions. These are rules that go in front of every chat. Models are trained to weigh them above your normal prompts, so put your most important rules there.

In Claude.ai, you do this with a project. A project keeps its instructions and its files for every chat inside it. Make one project per topic, for example one for your company paperwork with the documents attached. Then a quick question gets an answer that already knows the background. That is how you become an advanced Claude.ai user.

Still, Claude.ai is made for simple questions. It does not even show you how full the context is. Engineering work needs more control.

6. Engineers should work in Claude Code

Claude Code uses the same models as Claude.ai. Three things make it better for engineers.

Better instructions. It starts with a long hidden system instruction from Anthropic that teaches it to write good code.

More tools. It can read, write and edit the files on your drive, and run commands on your PC.

More control. You can shape it with commands, skills, subagents and rules. A basic user and an advanced user get very different results from the same tool. The difference is the setup.

My setup: install VS Code or Cursor (Cursor is a copy of VS Code with a look made for AI), then add the Claude Code extension. The editors are free. The extension puts Claude Code right next to your files and makes it easy to read. Type /context at any time to see how full the context is.

7. Skills: your procedures, loaded only when needed

A skill is a standard operating procedure written in a markdown file. If you can write a task as clear steps in plain English, it can be a skill. For example: load the model, check it for errors, fix the couplings, rename what needs renaming, run the tests. Write it once and the AI runs it for you from then on. You can also start it yourself with a slash command, like /analyze.

Skills load only when needed. Each skill has a one sentence description at the top, called front matter. Only that sentence sits in the context. When the AI sees that your task matches, it loads the full skill.

This is how my test agent, Testan, works. Its instructions run to hundreds of thousands of words: the full API, every HIL device, and more. A project that uses one device loads only that device's file. You can keep a terabyte of docs on disk and the context stays small.

8. Subagents: a fresh context for a separate job

A subagent is a second AI agent that the main one starts. It gets its own context window and its own instructions. It does one job and hands back the result.

Example: PDF documentation. AI cannot read PDFs well until they are converted. If the main agent does that work, a stack of manuals fills its context with work it never needs to remember. So I built a subagent called docs-to-text. Its only job is to turn every PDF, archive and picture into text, sorted by section, with one markdown menu on top. It does the heavy reading in its own context. The main agent reads only the clean text, and its own context stays clean.

Skills vs subagents

Both save context. They save it in different ways.

SkillSubagent
Where it runsInside the main agent’s contextIn its own, separate context
What the main agent seesThe full procedure, once it is neededOnly the result
Use it whenThe main agent needs the know-how for the taskThe job is separate, and you could hand it to another person
ExamplesBuild a full Modbus project from a short descriptionCode review, writing tests, model checks, docs-to-text

Rule of thumb: a skill is know-how the main agent should hold. A subagent is a job you can outsource. Code review is a good subagent. The reviewer does not need to know how the code was written. It only needs the rules for good code, and it reports back in one sentence.

9. Index files: the third option

Every skill costs one sentence of context, even when you never use it. With many skills, that adds up. Claude Code has a third option: folders of markdown docs with an index file.

For Testan, I downloaded the full API documentation as markdown, which is the cheapest format for AI to read. An index file lists what each doc covers. The index itself stays out of the context until the agent needs the API. Then it reads the index, and from the index it opens only the one file that matters.

This scales well for a team. Put converters in one folder with their own index, and machines in another. If a task has nothing to do with converters, nothing about converters enters the context. Not even one sentence. Claude Code finds files fast with its search tools, glob and grep. With subagents, skills and indexed docs, a team can give its agent a huge library of knowledge, and the agent uses only what each task needs.

10. Plan before you build

Claude Code has modes. In manual mode it asks you before every step. Other modes let it edit on its own. Each mode is just a hidden instruction added in front of your prompt.

Plan mode tells it to edit nothing yet. First it writes a plan in a markdown file and puts real effort into how it will do the job. You review the plan, say what you agree with and what you don't, and accept it. Only then does it start editing. Planning first works much better on real projects. Use it for anything bigger than a one line change.

I run Claude Code with permission prompts turned off, so I do not approve every step by hand. That saves a lot of time, but the agent can then do anything you can do on that PC. Only turn them off in a setup you trust.

11. Let AI write your skills, with rules

AI can turn your conversation into a skill file for you. Left alone, it writes long text that only another AI wants to read. Models are wordy by default. Then a colleague opens your project and cannot follow it.

Give it rules first: write at a 5th grade reading level, use short sentences, bullet lists, numbered lists, bold and clean formatting. You get a skill that works for the AI and that people can read too.

Four pieces of advice

Start now. AI is not going anywhere. When you need it, you want to be good at it already.

You are in the perfect spot. You already know how to do the work by hand. That is exactly what lets you judge what the AI gives you, instead of shipping AI slop.

Use Claude Code. VS Code or Cursor, plus the Claude Code extension. It is a quick install, and you will be amazed.

Build skills as you go. There is no big learning curve. Each time you repeat a task, turn it into a skill.

Want this set up for your whole team, with skills and agents built around your own work? That is what I install. It starts with the free AI audit. The hardware version of this class, with Altium examples, is AI for Hardware.