The No-Nonsense Engineer’s Guide to Claude Skills
Every engineer on your team is using AI. Most of them are also re-explaining the same context from scratch every single session. Their stack, their conventions, their preferences, their constraints. Every time. Gone the next day.
That’s the actual problem. And Agent Skills fix it.
What a skill actually is
A folder. A SKILL.md file inside it. That’s it.
The file has two parts: frontmatter that tells Claude when to load the skill, and a body that tells it what to do. Claude reads the name and description from every installed skill at startup, loads only the ones relevant to what you’re working on. You don’t prompt it to. It just happens.
---
name: post-mortem
description: Use when writing or reviewing post-mortems after production incidents
---
## Required sections
- Timeline of events (with timestamps)
- Contributing factors (not just root cause)
- Detection gap: why didn't we catch this sooner?
- Action items with owners and due dates
## What good root cause analysis looks like
Not: "the database ran out of connections"
Yes: "connection pool exhausted because the retry loop in payments-service had no backoff, introduced in commit abc123 on Tuesday"Every incident write-up session now starts with that context already loaded. You never re-explain it.
Why this matters more than prompting tips
Better prompts help once. Skills compound.
Your post-mortem format, your architecture review checklist, your on-call triage steps, your team’s definition of a P1 — that’s institutional knowledge that currently lives in a Confluence page nobody opens. A skill makes it load automatically in the tool your team uses every day.
The skill authoring best practices are worth reading. The short version: write the description like you’re telling a smart colleague exactly when to pull out a specific playbook. Vague descriptions are the primary reason skills fail to trigger when they should.
The format is now cross-platform
Anthropic created Agent Skills and released it as an open standard in December 2025. The spec lives in the agentskills GitHub org and has been adopted by Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenHands, and 20+ others. Same SKILL.md format, works everywhere.
The parallel to MCP is deliberate. Anthropic builds a spec that solves a real interoperability problem, releases it openly, and lets adoption do the work. Skills you write today aren’t locked to Claude. They travel with you across every agent that adopts the standard.
What Byte Sized Design is building on top of this
I’m putting together a set of skills for engineers. Free ones cover the workflows every senior engineer rebuilds constantly: structured debugging, architecture review, incident triage. Paid ones go deeper: post-mortem generation from raw timelines, system design critique grounded in real production tradeoffs, on-call runbook generation.
More details soon. If there’s a specific workflow you want to stop rebuilding from scratch, hit reply.
What to Take Back to Your System
Skills are the abstraction layer for your expertise. Stop re-explaining context. Encode it once in a SKILL.md and let it load automatically.
The description field is the trigger. Write it precisely: what the skill does and exactly when to use it. Vague descriptions mean the skill never loads when you need it.
The format is cross-platform. Skills work across Claude, Cursor, Gemini CLI, GitHub Copilot, and more. What you build today isn’t locked to one tool.
Institutional knowledge belongs in skills, not Confluence. Your team’s conventions, formats, and standards should load automatically, not sit in a doc nobody reads.