Public, reusable agent skills for UC San Diego and TritonAI workflows.
Created and maintained by David Balderston for UC San Diego.
An agent skill is a folder of instructions, references, scripts, and assets that an AI coding agent can load when a task matches the skill's trigger description.
This repository is intentionally lightweight. It is a skills library, not a dashboard, installer, generated catalog, or internal platform runbook.
UCSD-Skills-Library/
README.md
AGENTS.md
CONTRIBUTING.md
LICENSE
tritonai/
README.md
skill-name/
SKILL.md
LICENSE
references/
assets/
scripts/
community/
README.md
skill-name/
SKILL.md
LICENSE
references/
assets/
scripts/
tritonai/is for skills maintained by the TritonAI or UCSD AI Tools team. Only authors on the private AI team allowlist may contribute skills there.community/is for skills contributed by everyone else and reviewed before merge.- Each skill lives in its own folder.
- Each skill must have a
SKILL.mdentrypoint and aLICENSEmatching the repository MIT license.
Copy the individual skill folder into the skills directory used by your agent. Do not copy the tritonai/ or community/ wrapper folder.
Common locations:
| Use case | Put skill folders here |
|---|---|
| Global dotagents skills | ~/.agents/skills/ |
| Citizen-Developer repo-local skills | <Citizen-Developer repo>/.agents/skills/ |
| Managed TritonAI Harness install | ~/.agents/ucsd/skills/ |
For global dotagents-compatible setups:
mkdir -p ~/.agents/skills
cp -R tritonai/example-skill ~/.agents/skills/For the Citizen-Developer workspace:
cd "/path/to/Citizen-Developer"
mkdir -p .agents/skills
cp -R "/path/to/UCSD-Skills-Library/tritonai/example-skill" .agents/skills/For the managed TritonAI Harness install:
mkdir -p ~/.agents/ucsd/skills
cp -R tritonai/example-skill ~/.agents/ucsd/skills/After copying, the installed skill should look like this:
~/.agents/skills/example-skill/
SKILL.md
LICENSE
references/
assets/
scripts/
Every SKILL.md starts with YAML frontmatter:
---
name: example-skill
description: Use when an agent should do a specific workflow. Trigger on concrete user intents, keywords, file types, or slash commands.
---The frontmatter description should explain when the skill should be used. Keep it concrete so agents do not load the skill for unrelated work.
Keep frontmatter minimal. Public TritonAI skills require name: and description:; allowed-tools: is optional. Community skills must also include a maintainer: field naming the contributor, team, or organization responsible for the skill.
This public repository should only contain skills that are safe to publish openly.
Do not add:
- Secrets, tokens, API keys, private certificates, or credentials.
- Private UCSD infrastructure details or deployment procedures.
- Real student, patient, employee, customer, or operational data.
- Internal-only runbooks, escalation paths, or restricted service assumptions.
- Skills that send email, write to production systems, or use authentication without clear user confirmation and review.
Internal or restricted skills belong in the private UCSD-Skills-Library-Secure repository instead.
Some skills are maintained next to the code they describe in another
repository. external-skills.json lists them, and
.github/workflows/sync-external-skills.yml copies each one in every Monday or
on demand from the Actions tab, opening one pull request per skill when the
upstream copy changes. For example, tritonai/ucsd-decorator/ comes from
UCSD/decorator-kit.
Each entry looks like this:
{
"name": "ucsd-decorator",
"collection": "tritonai",
"repository": "UCSD/decorator-kit",
"path": "library/tritonai/ucsd-decorator",
"track": "latest-release",
"ref": "v2.5.0",
"commit": "063ecc6280a2a5fabe7121a837c0f0c05235c67f"
}track: latest-releasefollows the newest GitHub release.track: pinnedstays onrefuntil someone changes it.refandcommitrecord what was copied; the sync updates them.- The upstream repository
LICENSEmust match this repository'sLICENSEexactly. The sync refuses anything else so the copy keeps its original terms. - Do not edit synced folders by hand. The next sync overwrites them, and the preflight check warns about such edits. Request changes upstream instead.
To add an external skill, add an entry with ref set to the release to start
from, then run:
python3 scripts/sync_external_skills.py --skill <name>Commit the copied folder together with the updated manifest. Pull requests
opened with the default Actions token do not trigger the preflight workflow;
add a SKILLS_SYNC_TOKEN secret with contents and pull-request write access so
they do.
Pull requests are reviewed through three complementary layers:
- GitHub Actions runs lightweight preflight checks for contributor placement,
public-skill format, obvious leak patterns, and whitespace.
To avoid exposing private membership, public checks report maintainer
verification for
tritonai/changes without failing just because the private allowlist is unavailable. - CodeRabbit is configured by
.coderabbit.yamlfor AI review on each PR update, with emphasis on public-vs-secure repository fit. - The local Codex webhook reviewer in
docs/public-pr-review-service.mdruns through the Codex app server and posts public-skills review comments for each newly reviewed PR head SHA and issue update.
Contributions are welcome through pull requests, including Harness’s Share with UCSD action. See CONTRIBUTING.md for the expected skill layout, review rules, and public/private boundary.
This repository is available under the MIT License. See LICENSE. Each
individual skill folder also includes the license so it remains available when
that skill is copied or distributed on its own. Community contributors remain
credited through skill maintainer metadata and Git history.