Be Digital · Field notes

Why Claude “Skills” quietly change the economics of AI work

I went to a Claude Code bootcamp expecting to get better at prompting. The lesson that stuck was about something else: how to give an AI the context it needs without drowning it.

Progressive disclosure of context across three tiers Tier one, CLAUDE.md plus skill descriptions, is loaded every session and should be kept lean. Tier two, the full skill procedure, loads only when triggered. Tier three, scripts and resources, loads rarely. Skills cost almost nothing until used. Progressive disclosure: load the minimum What Claude holds in context, and when it pays for it ALWAYS LOADED — EVERY SESSION CLAUDE.md + skill descriptions Always paid · keep lean LOADED ON TRIGGER Full SKILL.md procedure Paid only when used LOADED AS NEEDED Scripts · docs · resources Paid rarely A library of 20 skills costs almost nothing — until one is needed. Always-on context is the tax. Skills move the rest off the meter.
The mental model: keep the always-on layer lean; let everything else load on demand.

The unlock: Skills and progressive disclosure

A Skill is just a folder with a SKILL.md file — a procedure you write once and reuse. The clever part is progressive disclosure: only the skill’s one-line description stays in context all the time; the full instructions load only when a task actually needs them. So you can keep a deep library of skills and pay almost nothing for them until one fires.

Contrast that with CLAUDE.md, which loads on every session. Every line there is a tax on every interaction — and it quietly dilutes the model’s attention, so the rules that matter get buried under the ones that don’t. That’s why the file has to stay lean: short isn’t a style preference, it’s how the model keeps following the important rules.

How I now set up a repo for AI work

Rules and pointers go in CLAUDE.md (lean, always on). Procedures become Skills (load on demand). Durable facts live once in a memory file (one home, no duplication).

Three take-aways I’m keeping

  1. Keep CLAUDE.md short. If a line wouldn’t matter for most sessions, it belongs in a Skill.
  2. One home per fact. Duplicated guidance costs tokens and drifts into contradictions.
  3. Write sharp, non-overlapping skill descriptions. That’s what decides whether the right one fires.

Net effect: fewer wasted back-and-forth turns, lower token spend, and an assistant that follows the rules that count. I’m still chewing on a few open questions — like how far to push repo “maps” — but this mental model alone was worth the trip.

Be Digital — notes from the bench. Further reading: The Complete Guide to Building Skills for Claude · Effective context engineering for AI agents

Deploying GenAI in a regulated environment?

See GenAI & AppSec advisory