AI Guides › Skills & Agents

The Real Cost Of A Marketplace Full Of Skills You Never Read

By Nigel Guy · 2 min read

A marketplace with hundreds of ready-made skills feels like pure upside — more capability, installed in seconds, no writing required. The cost doesn't show up at install time. It shows up later, quietly, as a library nobody fully understands, running actions nobody in the room actually reviewed.

The mechanism: the Install-Time Ledger

Before adding a skill from a marketplace, however well-reviewed, answer these four questions and keep the answers somewhere you'll see them again:

1. What does it actually do, in your own words? Not the marketplace description — read the skill itself and restate its behaviour. If you can't restate it after reading it, that's information too.
2. What can it touch? Read access, write access, external calls — list what it's capable of, not just what it's meant for. A description saying "summarises your inbox" and a skill that also drafts and sends replies are not the same risk.
3. Who else has vetted it, and how? A high install count tells you plenty of people ran it; it doesn't tell you any of them read it closely. Popularity is not the same evidence as review.
4. What happens if it's wrong once? Same question you'd ask of an agent's write access — apply it to every skill you add, not just the obviously risky ones, since low-stakes-looking skills are exactly the ones that get installed without a second look.

A library where every entry has answers to these four questions is an asset. A library where most entries don't is a liability wearing an asset's clothes.

What to skip

Skip installing "just to try it" and then leaving it in the library indefinitely — a trial you never followed up on is functionally identical to a blind install, just with extra steps. Skip trusting a skill more because it came from a well-known marketplace; a marketplace can vet for malware without vetting for whether the skill's actual judgement is any good at your specific task.

Guardrails

• Reading a skill once doesn't mean it stays trustworthy — a marketplace skill can be silently updated by its author, so "I read it in the past" is not the same claim as "I know what it currently does."
• This mechanism doesn't replace platform-level security review; it's the judgement layer on top of it, for the specific question of whether a skill fits your task and your risk tolerance.
• The size of a library is not itself a virtue. A smaller library of skills you've actually read beats a larger one you haven't, every time.

All 751 AI guides · JulieMango plans from £17/mo