All posts

The Install Said Success. My Coding Agent Still Couldn’t See the Skill.

4 min read
  • #technology
  • #artificial-intelligence
  • #developer-tools
  • #programming
  • #software-development

Why I built skill-land to install and verify agent skills where they are actually read

I ran an install command for an AI coding agent skill. The terminal printed a success message. The command exited with code 0.

Then I checked the directory Codex reads. It was empty.

The skill had been copied to ~/.agents/skills/skillmama. Codex was looking in ~/.codex/skills/. The installer had done something, reported success, and left the agent without the skill.

I reproduced this with skills@1.5.22 on August 8, 2026. The package has continued to change since then, so this is a report about that specific reproduction, not a claim that every current install fails. The larger lesson stayed with me: a successful write is not proof that the agent can use what was written.

Two different questions

Installing a skill raises two questions:

1. Did the files get copied?
2. Did they land somewhere the agent reads?

Most install feedback answers only the first one. That leaves the developer to open folders, guess at configuration paths, and figure out why an agent cannot see a skill that appears to be installed.

That gap is what I built skill-land to address. It installs an agent skill to the target agent’s read paths, then checks that the files landed there. If a check fails, it reports the failure and exits with a non-zero status.

For example, to install the SKILLmama skill for Codex and Antigravity:

npx skill-land Magithar/SKILLmama --for codex,antigravity

To check an existing install without writing anything:

npx skill-land Magithar/SKILLmama --for codex --verify

Verification means reading the files back

For each installed skill, skill-land checks whether the expected files exist, are non-empty, match the source byte for byte, and have the expected name in SKILL.md front matter.

That catches errors a directory listing can miss. Two copies can have the same size and different contents. A multi-file skill can also appear to install while referenced scripts or examples are missing. Reading the files back makes the check about the result, rather than the installer’s success message.

There is another level of verification too. Files being on disk does not prove the agent loaded them. The optional --discover command starts a fresh session for supported agents and checks whether the skill appears in the agent’s own listing. It is opt-in because Codex and Claude Code discovery can make an API call; other agents may report “not checked” when a reliable listing is unavailable.

Skills also deserve a security check

Agent skills are instructions that can influence tools with access to a developer’s files and accounts. Before installing a skill, it helps to know what it contains.

skill-land discloses things such as shell commands, network hosts, and credential references it finds. When NVIDIA’s SkillSpector is available, it also reports its findings. By default, those findings are shown to the user. --strict can refuse flagged skills.

The scan is a signal to inspect, not a guarantee that a skill is safe. Static scanners can flag defensive documentation as well as harmful instructions. I want the tool to show what it found without pretending a score can settle the decision for you.

A small tool for a quiet failure

The most frustrating bugs are sometimes the ones that look finished. The command ran. The output was green. The thing you needed still was not available.

I built skill-land around that gap: install a skill where an agent reads it, verify the files that arrived, and make failures visible. It can also audit a directory of skills you installed earlier, regardless of which tool copied them there.

The current npm release is 1.4.0. Try it with a skill source and a target agent:

npx skill-land <source> — for <agent>

Read the full usage and limitations on npm. If you have ever had an installer say “done” while your agent still could not find the skill, I’d be interested to hear what happened.

Also published on Medium. Read or respond there