Agent Skills following the open agentskills.io standard. Works with Claude Code, Claude.ai, OpenAI Codex, Cursor, Gemini CLI, GitHub Copilot, and other compatible agents.
| Skill | What it does |
|---|---|
learn-simply |
Explains any topic in plain, unambiguous language using Simplified Technical English rules |
knowledge-graph |
Maps a codebase — one page per feature — and tells you when a page has fallen behind the code |
product-videos |
Makes launch, tutorial, explainer and social videos of a software product, with voiceover, music and captions, showing only what the product really does |
Most AI explanations are fluent and unclear at the same time. They stack metaphors, use a fancy word where a plain one exists, introduce three terms in one sentence, and leave you re-reading the paragraph.
learn-simply applies the writing rules from ASD-STE100 (Simplified Technical English) — a controlled-language standard built so that technical text reads unambiguously for non-experts — to explanations. One idea per sentence. Active voice. Every term defined the moment it appears. One word for one thing, all the way through.
It does not dumb the content down. It removes the friction between the content and your head.
Without the skill:
JWTs are signed tokens that the server issues after authentication and that the client subsequently transmits on every request, enabling the server to validate identity statelessly without requiring a database lookup.
With the skill:
A JWT is a signed token. The server creates it after you log in. The client sends it back with every request. The signature lets the server confirm who you are without a database lookup.
Same information. One of them you read once.
- Adaptive strictness. Familiar topics get the STE spirit with room to breathe. Hard topics (distributed systems, cryptography, economics) get the rules enforced tightly — one concept at a time, nothing new introduced until the last thing landed.
- A four-layer structure. Core idea → an analogy with its breaking point stated → how it actually works → a short Q&A recap that doubles as a self-test.
- Honest analogies. Every analogy is checked for whether it would predict something false if you extended it. If it would, the skill tells you where it breaks.
- Flagged simplifications. If a simplification is technically a lie, it says so rather than letting you walk away with a wrong model.
- Layered for big topics. Ask it to teach you a whole field and it gives you a map plus the first sub-topic, not a compressed dump.
Claude Code (as a plugin):
/plugin marketplace add CodewithVeek/skills
/plugin install learn-simply@codewithveek
Any agent (via the skills CLI):
npx skills add CodewithVeek/skillsManually — copy the skill folder into your agent's skills directory:
git clone https://github.com/CodewithVeek/skills.git
cp -r skills/skills/learn-simply ~/.claude/skills/ # Claude Code
cp -r skills/skills/learn-simply ~/.codex/skills/ # OpenAI CodexClaude.ai — zip the skills/learn-simply folder and upload it under Settings → Capabilities → Skills.
You mostly don't have to. The skill triggers on its own when you ask to learn or understand something:
explain how database indexes actually work
teach me consensus algorithms
what is a CDN
ELI5 how interest rates move currencies
To force it in Claude Code: /learn-simply:learn-simply
skills/learn-simply/
├── SKILL.md # Rules, structure, and adaptive strictness
└── references/
└── ste-rules.md # Full rule list with before/after rewrites,
# loaded on demand for strict-mode explanations
The reference file is loaded only when the agent needs it, so the skill costs almost nothing in context until it's actually working.
Documentation rots quietly. Nothing tells you which page stopped being true, so people stop trusting all of it, and then nobody updates any of it.
knowledge-graph keeps a codebase's knowledge as a folder of small pages — one per feature — with an index that sets the reading order, previous/next links at the bottom of every page, and one piece of metadata per page: the source it describes.
<!-- kg:covers
src/lib/webhook-handler.ts
src/api/webhook/
-->That comment stays invisible when the page renders, and it is what lets tooling answer the question documentation can never answer about itself: what is out of date right now?
One file, no dependencies, Node 18+.
| Command | Answers |
|---|---|
node $KG check |
Is every page in the index and every indexed page real, do all links resolve, do the covered paths exist, are the footers current? |
node $KG stale |
Which pages have had commits to their code since the page itself last changed — and which commits were they? |
node $KG nav --write |
Rebuild every previous/next footer from the index's order |
node $KG map --gaps src |
What does each page cover — and which code files does nothing cover? |
- Reads before exploring. In a repo that has a graph, the agent reads the index and the pages for the area first. Cheaper than re-deriving the structure every session, and it carries the lessons that are not in the code.
- Updates as part of the work. At the end of any change: which pages does this touch, edit them, regenerate the navigation, run the check. Not a follow-up task that never happens.
- Verifies before it writes. Claims get checked against the source, never carried over from another document. For a batch of pages there is a deep pass — subagents fact-checking every claim adversarially, plus one hunting for what is missing. On the project this skill came out of, that pass found 41 errors in 18 freshly written pages.
- Fixes the code when the code is the problem. A verification pass that only ever edits prose is missing half of what it finds: a log level that contradicts what the docs promise is a bug, not a documentation error.
- Bootstraps a repo that has none. Survey, agree the page list, write the index first, then fill it in.
Claude Code (as a plugin):
/plugin marketplace add CodewithVeek/skills
/plugin install knowledge-graph@codewithveek
Any agent (via the skills CLI):
npx skills add CodewithVeek/skillsManually:
git clone https://github.com/CodewithVeek/skills.git
cp -r skills/skills/knowledge-graph ~/.claude/skills/ # Claude Code
cp -r skills/skills/knowledge-graph ~/.codex/skills/ # OpenAI Codexdocument this codebase
update the knowledge graph
is the documentation still accurate?
what has gone stale?
where does payment capture actually happen?
To force it in Claude Code: /knowledge-graph:knowledge-graph
skills/knowledge-graph/
├── SKILL.md # When to read it, when to update it, the rules
├── references/
│ ├── page-contract.md # Page template, naming, what does not belong
│ ├── bootstrap.md # Starting a graph in a repo that has none
│ └── deep-verification.md # Fact-checking every claim against the source
└── scripts/
└── kg.mjs # check · stale · nav · map
The reference files load only when that part of the job comes up, so the skill stays cheap until it is actually working.
Launch videos like Linear's or Raycast's are mostly code now: the product's screens, a camera that zooms in on the action, a cursor that types and clicks, words timed to a voice. product-videos turns that into a repeatable process with Remotion.
It rebuilds the product's real screens as React components from its code and screenshots, so numbers count up and the camera can zoom without blur. And it holds one rule above the rest: everything shown exists in the product. Motion is free; invented buttons, toasts and live updates are not.
- Launch videos: problem → "Introducing …" → pitch → product moments → close
- Tutorials and walkthroughs: step by step, with the app's real labels and confirmations
- Explainers: diagram-led "how it works" videos
- Social cutdowns: 9:16 versions from the same scenes, with captions burned in
- Five visual styles, each with a recommended default per video type:
- Studio: dark stage, the real app in a window, a camera and cursor (tutorials, onboarding)
- Editorial: warm paper, one big rolling number per scene, floating cards, focus pulls, brand-colour circle reveals (launches, feature drops, explainers)
- Graphic: white grid, one loud brand colour, pixel motifs, colour slabs, logo swirls and orbits (API and infrastructure products)
- Cinematic: near-black, soft glow, large light type, slow (premium brands, teasers)
- Showcase: white, gradient accent words that drop in along curves, the product working live in a browser, cards lifting out (AI, agent and automation products)
- Motion measured from real launches: every frame of four professional launch videos analysed with optical flow, each animation fitted to Remotion's own easings and verified by re-rendering. The kit ships the curves as presets by role (
screenPan,settle,textRise…), four measured libraries, and the tooling to measure your own reference video and score a render against it - Modern type, bundled: Inter Display for headlines, EB Garamond with italic accents for editorial, Instrument Serif italic accents for cinematic, all from npm, so renders work offline
- Music and sound that hit: a 128 BPM pop-house track and a calm 100 BPM beat, mixed to match measured launch tracks and normalised to −13 LUFS, plus a crisp click, a swept whoosh and a pop, all synthesised (no licence needed)
- Plus voiceover (local Kokoro, or ElevenLabs),
.vttcaptions, loudness at -16 LUFS, and thumbnails embedded as cover art
- Recommends a style, and says why. It suggests the visual style that fits your video type, product and channel, explains the choice, lists the alternatives, and lets you pick.
- Asks first, then storyboards. Audience, video type, visual style, length and shape, whether you bring a script, voice and music. Then it shows a scene-by-scene storyboard before building.
- Checks every claim. Your script's lines are checked against the product, and overclaims get a precise rewrite ("across borders", not "globally").
- Narration drives timing. Each voice line is pinned to the frame of its action, and scenes grow to fit.
- Verifies what it can't watch. Frames are checked as stills, and sound is checked by measuring it. It tells you what's left to judge by eye and ear.
- Ships a starter kit. A tested Remotion project with the camera and anchor system, cursor, cards, thumbnails, vertical layout, and scripts for voice, music, captions, loudness and an audio report.
- Matches a reference's motion, measurably. Give it a reference video and it can measure how it moves, reuse those curves, and score its own render against the reference out of 100.
Claude Code (as a plugin):
/plugin marketplace add CodewithVeek/skills
/plugin install product-videos@codewithveek
Any agent (via the skills CLI):
npx skills add CodewithVeek/skillsManually:
git clone https://github.com/CodewithVeek/skills.git
cp -r skills/skills/product-videos ~/.claude/skills/ # Claude Code
cp -r skills/skills/product-videos ~/.codex/skills/ # OpenAI CodexNeeds Node 22+. Voiceover adds kokoro-js (about 500 MB, installed once).
make a launch video for my app, it's running on localhost:3000
turn the onboarding flow into a 60-second tutorial with a voiceover
make a 9:16 version of the launch video for Reels
make a 30-second launch video for our API in the graphic style
make a LinkedIn launch video like this one (attach a reference video)
make a showcase-style launch video for our AI agent
measure how this reference video moves and use that timing
add music and click sounds to the videos
write the VTT captions and a thumbnail for each video
To force it in Claude Code: /product-videos:product-videos
skills/product-videos/
├── SKILL.md # Intake questions, the workflow, the rules, delivery
├── references/
│ ├── styles.md # Launch, tutorial, explainer, teaser, cutdown… structures and pacing
│ ├── visual-styles.md # The five looks, which to recommend for which video, and why
│ ├── motion-data.md # Motion measured from real launches: the numbers, presets by role, measuring a reference
│ ├── scriptwriting.md # Word budget, headlines, narration, claim audit, pronunciation
│ ├── audio.md # Voice, music, effects, levels, licences, ElevenLabs MCP
│ ├── accuracy.md # What may be shown, how to verify it
│ ├── kit.md # The starter project: looks, transitions, motion blocks, camera, anchors, scenes
│ ├── social.md # Platform shapes, safe areas, captions, thumbnails
│ └── pitfalls.md # Bugs met along the way, and their fixes
├── assets/starter/ # A working Remotion project with Studio, Editorial, Graphic and Showcase example videos
└── assets/motion-lab/ # Measure a reference video, fit Remotion presets, score renders against it
Issues and pull requests welcome. If an explanation the skill produced was still confusing, open an issue with the topic and the output — that's the most useful bug report for a skill like this.
MIT — see LICENSE.