"References" hallucinated?

[zhsloe]

Multiple skills mention non-existing .md files in the non-existing "references" folder. Am I missing something or did the AI that generated the skills just hallucinate about those references?

[mrfeixiang]

Short answer: you are not missing anything. Those references are very likely hallucinated by the AI that generated the skills.

What is happening

In the LabClaw repository many skills follow a pattern like:

skill.yaml
prompt.md
references/*.md

The problem is that the repository often contains only the skill definition and prompt, while the references directory was never committed. When the skills were generated automatically, the model added “references” files that sounded reasonable, but they were never actually created.

So when the runtime loads the skills it sees something like:

references:

references/fda_database.md

references/pubmed_api.md

but the folder

references/

does not exist in the repository, which produces the warning you saw.

Why this happens

Most AI-generated skill packs are produced with a template that includes optional documentation sections:

references
examples
datasets

The generator fills the YAML assuming those files exist, but the generation pipeline does not actually create them. This is a common artifact in many agent tool repositories.

Does it break anything

Usually no.

OpenClaw only needs:

skill.yaml
prompt.md
(optional) code script

The references section is typically just extra context for the model. If the files are missing, the skill itself still works.

Three practical ways to handle it

Option 1 (recommended): ignore the references field

If the skill works, you can simply ignore those warnings. Many LabClaw users run it this way.

Option 2: remove the references lines

Open the skill YAML and delete lines like

references:

references/xxx.md

This removes the warning completely.

Option 3: create an empty references folder

For example

mkdir references
touch references/placeholder.md

Some loaders stop warning once the directory exists.

A deeper point

This is actually a good reminder that many “AI tool libraries” are still semi-synthetic. The useful part is the workflow logic, not the documentation artifacts. When deploying these systems in a research environment, it is often better to curate a smaller set of skills and maintain them manually.

[zhsloe]

Well this makes me actually question the quality of this whole skill set, as it is supposed to be carefully curated — the devs say it themselves in the readme. Seeing these hallucinations is a very bad sign for me personally.

Also would appreciate it if I got answers form a real human and NOT from another AI.

[zhsloe]

Actually, never mind my question - every skill that is supposed to have references comes from K-Dense Inc. and there are 115 of them:

$ grep -rl "references/" | xargs awk -F': ' '/skill-author/ {print $2}' | uniq -c
    115 K-Dense Inc.

And the skills in the original K-Dense repository do have a references folder where required. I guess the only question that stays is: why does this repo not include everything that comes with the original skills?