This week we're talking Skills: a feature many experienced devs seemingly love to dismiss as "just markdown files."
(Spoiler: they're wrong, and it's costing them real leverage)
In today’s edition talk about why skills are actually folders (not files), a round-up of the latest agent updates worth knowing, and a prompt to help you turn one of your own messy workflows into a reusable skill.
Let's get into it 👇️
I keep seeing the same bad take from many experienced devs that I follow.
“Skills? Don't need 'em.”
“I've been coding 15 years. I just write good prompts.”
Or my favorite: “they're just markdown files.”
And fair enough, you can do the work without them.
But that doesn't mean you should.
It's holding you back from what is, in my opinion, the best framework there is for getting the most out of an LLM.
Skills aren't markdown files. Skills are folders.
Inside those folders you can host subfolders, scripts, templates, even sub-skills. The SKILL.md is just the README for the agent. It tells the agent how to explore the folder, what to use, and when.
Calling a skill "just a markdown file" is like calling a repo "just a README."
Your coding agent will ONLY load into context what it actually needs, one step at a time. HIGHLY token efficient.
With this approach you can build powerful, repeatable automations:
Scripts handle the work that needs to be exact: API calls, file operations, formatting.
LLMs handle the work that benefits from flexibility: writing, research, synthesis.
You're not choosing between rigid automation and flexible AI, because you're combining both in a single orchestrated system.
Let me walk you through a skill-based automation that I built to run an entire pipeline for a financial newsletter:
I hand it three URLs.
It runs a script to pull seven days of price data.
Then it spins up three subagents in parallel, each calling a blog-writing skill.
They scrape the stories, run adjacent web searches for context, write the articles against a template, generate cover images via another script, and post everything to WordPress.
Back at the parent level, the agent assembles a newsletter from the three articles and outputs an inbox-ready HTML email.
The final tally for the whole automation: 10 skills, 3 subagents in parallel, 9 scripts, 4 templates.
One command kicks it off. 100% repeatable & reliable.
Every single run produces the same caliber of output.
The key concept here is the nesting:
A skill calls a subagent, and that subagent calls another skill. Each layer is self-contained but designed to plug into the layer above it.
You can think of it like functions in code. Each skill has clear inputs/outputs, and, just like functions, you can compose them into bigger systems.
The devs who understand this are building systems. The devs still dismissing skills as "just markdown files" are doing the work by hand.
(And that's fine, good for you! But you're not special for it either)
Anything you find yourself doing more than once can be packaged into a skill and run the same way every time.
Take a workflow you do over and over by hand and have Claude scaffold it into a real skill: folder, SKILL.md, scripts, and templates included.
Drop this into Claude Code (or any agent that supports skills) and watch it stand the whole thing up for you.
Paste it in, swap in your workflow, and run it 👇️
You are going to help me convert a repetitive workflow I do by hand into a Claude skill: a self-contained folder with a SKILL.md, scripts, and templates, so I can run it the same way every time.
WORKFLOW I WANT TO AUTOMATE:
<describe the workflow in plain English here. What triggers it, what inputs you give, what output you expect.>
Do this in order:
1. Ask me up to 5 short clarifying questions, but only if the workflow above is genuinely ambiguous. Otherwise skip straight to step 2.
2. Map the workflow into discrete steps. For each step, decide whether it should be a deterministic script (API call, file op, formatting, scraping a known structure) or an LLM step (writing, research, synthesis, judgment). Be explicit about which is which.
3. Propose a skill folder layout. Show me the tree: SKILL.md at the root, a scripts/ folder for the deterministic stuff, a templates/ folder for any reusable output shapes, and any sub-skills if the workflow is complex enough to nest. Name everything.
4. Write the SKILL.md. It should follow progressive disclosure: short instructions at the top, details only as needed, and clear pointers to which file in the folder to load and when. Include the description frontmatter.
5. Stub out each script with a docstring describing inputs, outputs, and what it does. Don't write the full code yet. Just the contracts.
6. Stub out each template with comments showing the slots that get filled in.
7. End with a one-paragraph "how to run this" note for me.
Constraints:
- Keep the SKILL.md under 200 lines. If it's longer, you're putting too much in the README and not enough in the folder.
- Anything that touches an external API, the file system, or a fixed data shape goes in a script, not a prompt.
- Don't invent steps I didn't describe. If something's missing, flag it instead of guessing. 
Your roundup of frontier-lab model releases, updates, and the moves worth keeping an eye on.
Claude Design. Anthropic shipped a new generative tool that builds full presentations, websites, brand videos, landing pages, and apps from a single text prompt. If you ship creative or marketing assets regularly, this collapses an afternoon's work into a prompt (Anthropic).
OpenAI's gpt-realtime-1.5. Low-latency voice API for interactive apps. Easiest path right now to ship a voice agent that doesn't feel laggy (OpenAI).
ChatGPT Images 2.0 (gpt-image-2). OpenAI's new image model thinks before it draws, holds character continuity across up to 8 images per prompt, and renders multilingual text accurately at up to 2K (4K via API). Replaces DALL-E across ChatGPT and the API. Currently #1 on Image Arena (OpenAI).
Bugcrawl in Claude Code. Anthropic's new feature scans your repo for bugs using 10 parallel agents. If you're on Teams or Enterprise, drop it on a stale codebase and see what surfaces (@WesRoth).
Advertise with Build with AI
Get in front of an audience of professionals using AI day-to-day: founders, engineers, operators, and product builders.
Interested in advertising? Respond to this email for rates and details.
Until next time,
William Ryan
Editor-in-Chief @ Build with AI
PS: Follow me on X for daily updates and AI workflows.