Спеціально для каналу Сашко пише код
Chapter 08 · For developers starting out

Skills are a real
engineering tool.

Not a magic productivity button. If you're just getting into AI coding agents, the honest default is start with nothing. When you catch yourself correcting the same mistake three or more times, that's when a rule earns its place.

The starter playbook

Pick one tool. Learn its skill system deeply.

Before trying to master all of them.

🅒

If you use Claude

Start with a short CLAUDE.md and add skills only for repeatable workflows.

🅒

If you use Cursor

Start with a couple of auto-attach rules scoped by file glob — these have near-zero token cost when you're editing unrelated code.

🅒

If you use Copilot

Write one copilot-instructions.md under 1,000 lines. If you use multiple tools, write an AGENTS.md at your repo root.

The three principles that survive

What actually matters after all the hype.

01

Descriptions are the single most important field

Write them like search queries. Second, keep each skill single-purpose and under ~40 lines of body content for most cases; reserve length for genuine reference material that lives in linked files.

02

Prune quarterly

Delete anything the model already does right, delete anything stale, delete anything contradictory. The patch spiral is real — Alex McFarland's widely-shared Substack: "Six rounds of fixes. Each round added more instructions."

03

Watch your token counts

Cursor shows them. Claude Code shows them. The Chroma context-rot research is clear: a rule that pushes you past 25% context fill is probably hurting output quality, not helping it. Add rules reluctantly and delete them aggressively.

By the numbers

What survives when the hype doesn't.

~100
tokens per skill idle
<5K
tokens when triggered
<60
lines per CLAUDE.md (recommended)
~1K
lines before Copilot degrades