The trouble with Agent skills isn't writing them—it's deleting them
Behind GitHub's Agent skills guide, the replies exposed a messier problem: skills expire, load incorrectly, and lack versions and tests. One person cut onboarding from 9 hours to 3.5 hours, while someone else audited 61 entries that had never been written to the record.
GitHub posted a tweet introducing Agent skills: save repeatable instructions so both the agent and you can reuse them. The tweet included a practical guide, linked here, with the original tweet here.
That description sounds like a prompt favorites folder. The real problem is that favorites folders rot.
First, solve repeated explanations
Commenter Jon Cheney said that previously every session required retyping instructions, but now instructions are centralized in one place and loaded only when a task needs them. Andrew put it more directly: write once, and every agent can get it.
GitHub's tweet also stated the value clearly: capture repeatable instructions. The goal is to stop the agent from starting from scratch every time.
The trouble starts when skills multiply
Javi mentioned a very concrete point: a skill is only useful if it is loaded correctly. After dozens of skills, choosing which one to load becomes a problem in itself. His approach is to record which skill was loaded on each run, so you can see when the wrong one was chosen.
ShadowAguy said skills only trigger when the description matches, and eventually he started writing regexes. Buswe reminded people that description is the basis for the agent deciding whether to load a skill; the body only comes in after a match. If the description is written wrong, the skill effectively doesn't exist.
Expiration is another pitfall. LaunchKit is still solving the problem of skills going stale after rules change. MartinOL was blunter: old instructions that no one rereads will lead the agent in the wrong direction. Krrish's approach is to have agents start deleting old notes. A skill library needs pruning, or it becomes technical debt.
Max Petrusenko did a synchronization audit and found 61 lines that had never been written, and the skill sets loaded across the fleet were inconsistent. This problem isn't flashy, but it is deadly.
Versions, tests, permissions
Tanisha Pritha asked whether reusable institutional knowledge should be version-managed like normal code, or just thrown into a folder. Yuna Han's answer is close to practice: a versioned SKILL.md lets the team decide what enters the repo, rather than keeping it in four people's heads.
Validation is harder than writing skills. Amy said the smart part is giving an independent verifier only the request and diff, so it doesn't grade its own homework. Nikolay Rodionov said that validating with hundreds of combinations of harness, model, and environment is usually the hardest. Bhurva drew a distinction: instructions portable and behaviour portable are two different release notes; a shared test set should cover the same task, the same expected evidence, and also when to stop.
Permissions are also unavoidable. 波若波罗蜜 mentioned that even if a skill loads successfully, it may still be unavailable in another client, so you can add a capability preflight that reports missing tools before the workflow starts.
Ignacio Ferro gave a small example. His book-conversion skill specifies that manuscript text goes in Markdown and layout goes in config. The agent still occasionally stuffs chapter titles into config, so the CLI checks with postext check --strict, exits with code 3, and the agent fixes it first before a human looks.
An example that adds up
Sam Woods said the first skill he would write is onboarding. A virtual assistant (VA) he worked with put three prompts in one workspace; one of them turns a new client's intake call into a usable SOP. According to her own logs, onboarding a single client dropped from about 9 hours to 3.5 hours, and she took on five more clients.
There's another easily overlooked point: doanbactam said Copilot also reads .claude/skills, so one skills folder can serve two agents without duplication.
Agent skills look like prompt management, but they are actually more like releasing software. You need versions, tests, load logs, permission checks, and cleanup mechanisms. Writing your first skill can start with a process like onboarding, where repetition is highest and the payoff can be calculated. Don't rush to pile up dozens; build deletion and logging first. If a skill library isn't pruned, it will eventually become a source of noise for the agent.
发布时间: 2026-11-17 15:31