Skills and the compounding loop
Procedural memory - the fourth point on the memory spectrum - is what turns a system that answers well *today* into one that gets measurably better at doing the same job *next time*, without retraining the model. Agent Skills are how you write procedural memory down.
What a skill actually is
A skill is a folder: a `SKILL.md` file (a name, a description of when to use it, and step-by-step instructions) plus, optionally, scripts and reference files the agent can run or read. Nothing exotic. It is a well-organised, well-labelled piece of your own knowledge that the agent loads only when the task matches the description.
The compounding loop, in four steps
- 1
Do a task manually once (or twice) with the agent
Notice the pattern: the same three-step process for onboarding a client, the same debugging checklist, the same report format.
- 2
Write it down as a skill
Capture the steps, the gotchas, and any reusable script, exactly as you would explain it to a competent new hire.
- 3
The agent picks it up automatically next time
A well-written skill description means the agent recognises "this looks like that job" and loads the instructions without you re-explaining anything.
- 4
Refine the skill when it misfires
Every time the agent gets the task slightly wrong, that is a signal to sharpen the skill file, not to re-teach the agent from scratch in chat next time.
Ready when
Why this compounds
Each skill you write makes the next similar task faster and more reliable, for you and for anyone else using the same agent. Ten well-written skills is a meaningfully more capable system than the same model with zero skills, even though the underlying model has not changed at all.
Skills vs the rest of memory
Semantic memory holds facts ("the client's renewal date is March"). Episodic memory holds events ("last Tuesday's deploy broke checkout"). Skills hold how, the repeatable process, independent of any single fact or event. That is why skills belong in version control next to your code, not in a database row.
Watch out
Common mistakes
- Writing a skill so broad ("handle customer stuff") that the agent cannot tell when it applies.
- Never revisiting a skill after it misfires, so the same mistake repeats indefinitely.
- Keeping tribal knowledge in your head instead of a skill file, so it never compounds for anyone but you.