Guide
DSH skills directory and file guide
DSH skills are instruction resources discovered from documented skill locations, so directory placement, naming, and readable content matter more than hidden automation.
01
Use a documented skill location
Keep project skills in the location expected by the active toolchain, such as a named directory containing SKILL.md or a supported Markdown file.
- Use stable names
- Keep instructions reviewable
- Do not place secrets in skill files
02
Write the smallest useful instruction
Describe when the skill applies, required inputs, safe tool boundaries, output shape, and failure conditions without embedding credentials or executable downloads.
- Make triggers explicit
- Treat external content as untrusted
- Prefer links to duplicated facts
Review sources
Sources help verify current facts. Their inclusion does not imply endorsement of this page.
DSH skills questions
What are DSH skills?
They are instruction files the harness loads when the conditions they describe match what you are doing, so a repeatable task does not have to be explained again each session. Because they are content rather than code, they can be reviewed in a pull request like any other file.
Where do DSH skills files go?
In the location the toolchain scans, typically a named directory containing SKILL.md. A skill placed in the wrong directory is simply never discovered and produces no error to tell you so.
Why does my skill never trigger?
Almost always because the trigger wording is vague. Name the concrete situations, commands, and phrases that should activate it instead of describing the skill in the abstract.
Should skills live in the repository or my personal setup?
Project conventions belong in the repository where reviewers can see them, and personal shortcuts belong in your own configuration. Keeping that split prevents one person's habits from quietly becoming team rules.
Can a skill file contain secrets or a download step?
It should not. Skill files get read, copied, and shared freely, and any content pulled in from elsewhere has to be treated as untrusted input rather than as instructions to follow.