claude-skills
A working set of 26 Claude Code skills, used daily on real projects. Code review, shipping, marketing assets, device control, and agent orchestration.
These are not demos. Every one of them started as a task that went wrong, or went slowly, enough times to be worth writing down.
What a skill is
A skill is a folder with a SKILL.md inside it. The file carries YAML frontmatter
with a name and a description, then plain Markdown instructions.
---
name: blindspot
description: Use before modifying unfamiliar or high-stakes code...
---
# Blindspot
...instructions Claude follows...
Claude Code reads the description of every installed skill at session start, and
loads the full body only when the description matches what you are doing. So a skill
costs almost nothing until it fires. You can also invoke one by hand as /blindspot.
Install
Skills live in ~/.claude/skills/. Take all of them:
git clone https://github.com/ucsandman/claude-skills.git ~/.claude/skills
Or take one:
git clone https://github.com/ucsandman/claude-skills.git /tmp/claude-skills
cp -r /tmp/claude-skills/blindspot ~/.claude/skills/
Start a new Claude Code session and the skills are live. Run /blindspot to check.
To scope a skill to one project instead of your whole machine, put it in
.claude/skills/ inside that repo.
The skills
Thinking and review
| Skill | What it does |
|---|---|
interview |
Interrogates a vague request before any code is written. For "make it better" and "add auth". |
blindspot |
Surfaces what you are missing before you touch unfamiliar or high-stakes code. |
adversarial-review |
Parallel review agents find issues, then a skeptic that defaults to REFUTED tries to kill each finding. Only survivors are reported. |
de-vibe |
Finds and fixes the tells that software was AI-generated, then gives it a small distinctive identity. |
show-me |
Explains the current topic visually: pseudocode, call trees, file trees, Mermaid, diffs, or one focused HTML page. From humanlayer/skills, MIT. |
frontend-verify |
Drives the app after a UI change. Reads console and network errors first, writes full page state to disk, and only pulls a screenshot into context when a route actually fails. |
Shipping
| Skill | What it does |
|---|---|
ship |
The end-of-work ritual. Docs, commit, push, and a security spot-check that blocks the ship like a failed test. |
wrap |
Session handoff. Writes durable notes and a resume prompt before a context or rate limit hits. |
harness-health |
Read-only check of the local Claude Code setup. Finds dead hook paths, orphaned hook scripts, and enabled plugins with no cache. |
fable-gpt |
Hands heavy implementation and debugging to Codex, and orchestrates Claude plus Codex in one session. |
Marketing and launch
| Skill | What it does |
|---|---|
marketing-studio |
Shared background for every asset skill below. Remotion, Blender, and Playwright pipeline. |
logo-reveal |
Logo animation and brand intro video. |
social-clip |
Short feature-announcement clip for X, LinkedIn, or TikTok. |
product-demo |
Screen recording with zooms and a real cursor. |
launch-video |
30 to 90 second announcement composing demo footage, logo, and copy. |
og-assets |
OG images, link previews, animated OG loops, and README GIFs. |
audio-track |
Generated music, voiceover, and narration. |
marketing |
Runs the whole asset suite for a product in one pass. |
launch |
Launch copywriting and strategy, including the virality and positioning references. |
ship-it |
Renders the assets, then posts the launch with videos attached. |
Agents and orchestration
| Skill | What it does |
|---|---|
team |
Fans one task across two agents. Picks a lead, delegates bounded subtasks, logs every exchange to a ledger, and reports back. |
meditate |
A nightly reflection loop. Observations have to earn their way up a ladder from note, to memory, to standing rule. Most nights promote nothing, which is the point. |
dashclaw-governance |
Governance protocol for a supervised agent. Risk thresholds, approval waits, and action recording. |
Devices
| Skill | What it does |
|---|---|
phone |
Drives a real iPhone over USB. Send and read messages, open apps, tap, scroll, and OCR the screen. |
phone-gotchas |
The coordinate traps and silent failures that iPhone automation hits. Required reading before the first tap. |
deskclaw |
A read-only view of the Windows desktop. Lists windows, dumps a window's accessibility tree, and takes screenshots. |
Personal
| Skill | What it does |
|---|---|
wes-voice |
Rewrites a draft so it sounds like a person wrote it quickly, not like marketing copy. |
Layout
claude-skills/
├── LICENSE
├── README.md
└── <skill-name>/
├── SKILL.md # required: frontmatter + instructions
├── references/ # optional: loaded on demand, keeps SKILL.md small
├── scripts/ # optional: executable helpers
└── evals/ # optional: test cases for the skill
The references/ split matters. SKILL.md stays short so it is cheap to load, and
the long material sits in a reference file that Claude reads only when it needs it.
Before you clone
A few of these skills call local tooling by path, such as a rendering pipeline or a
message ledger, and expect it to exist. Read the SKILL.md before running one and
adjust the paths for your machine. The reasoning in each skill is portable even where
the paths are not.
phone and deskclaw are Windows plus iPhone specific. Everything in Thinking and
review is plain Claude Code and runs anywhere.
Third-party skills
Some skills installed alongside these are other people's work under their own licenses. They are deliberately not redistributed here. Install them from source:
| Skill | Upstream |
|---|---|
claude-api |
anthropics/skills (Apache 2.0) |
six-hats, lateral, inversion |
danium/lateral-thinking |
polish |
pbakaus/impeccable |
agent-browser |
vercel-labs/agent-browser |
Everything tracked in this repository is original work.
Contributing
Issues and pull requests are welcome, especially bug reports on a skill that misfires or fails to trigger when it should. If you fork a skill and make it better, a link back is appreciated but not required.
License
MIT. See LICENSE.
Support
If my tools save you time, you can support my work here:
No comments yet
Be the first to share your take.