LLM SkillsCreate Skills for your agents
Create LLM Skills for your agents.
SkillDiscs turns real sources into clean data infrastructure for agents. Start with a GitHub repository, book, research paper, or web article; parse and vectorize it as a Disk; then use it as a knowledge base, MCP source, REST API, or Claude Skill.
Supported formats and inputs
PDF filesYouTube videosDOCX · PPTXText & notes
app.skilldiscs.com / library
Human · reading
FyFeynman explanation100%
RqRecall quiz · 8/10—
CmConcept map—
RsRSVP focus mode—
SKILL DISC
AAgent · querying
RsREST searchsyncing
McMCP retrieval—
SkClaude Skill—
ViVector index—
LIVE
Definition
What is an LLM Skill?
An LLM Skill is a self-contained folder an AI agent loads only when a task calls for it: a SKILL.md file holding a name, a description, trigger phrases and the procedure itself, plus any reference files it points to. The description is what the model reads to decide whether to open the Skill at all — so a Skill is less a document than a procedure with a trigger attached.
SkillDiscs builds one from a source you already trust — a repository, a PDF, a book, a talk — by extracting the procedure it describes and packaging it in that format, with references back to where each part came from.
Architecture for AI
Clean data infrastructure for better AI output.
Agents need relevant, structured data instead of raw PDFs and prompt fragments. SkillDiscs parses, curates, vectorizes, and exposes each Disk through retrieval, API, MCP, or Skill export.
Knowledge graphconcepts, links, hierarchy
Markdown structureclean sections for prompts
JSON elementsheadings, tables, figures
Retrieval + accessembeddings, pgvector, API/MCP
Docs are live
The complete API and MCP reference lives at docs.skilldiscs.com.
Use the docs for endpoint schemas, MCP tools, redaction behavior, authentication, and copy-paste examples for agents.
Agents need clean data, not longer prompts.
Most agent workflows run on raw HTML, messy PDFs, or copied prompt fragments. That increases hallucination risk. SkillDiscs prepares the source first: parsing, structure, concepts, JSON/Markdown, vector index, retrieval, and Skill instructions in one Disk.
You curate the source. SkillDiscs turns it into usable infrastructure for your agent.
Five steps
How SkillDiscs creates an LLM Skill
From a source you trust to a folder your agent can load.
- 01
Add the source as a Disk
Paste a URL, upload a PDF or Office file, drop in a YouTube link, or connect a GitHub repository. SkillDiscs parses it into sections, key points and concepts, keeping the original attached.
- 02
SkillDiscs checks the Disk is procedural
A classifier decides whether the content actually describes a repeatable procedure. Reference material that does not is flagged as such rather than forced into a Skill that would never fire correctly.
- 03
Extract the procedure
Steps, preconditions and the passages they rest on are pulled out of the Disk. Each part keeps a reference to the section of the source it came from.
- 04
Generate the Skill package
You get a SKILL.md with a name, a description and trigger phrases, plus the reference files the procedure points at — the folder layout Claude and Claude Code load directly.
- 05
Evaluate the trigger, then download
The generated Skill is scored against on-target prompts (it should fire) and off-target prompts (it should stay quiet), so you see how reliably it activates before you ship it. Download the package and drop it into your agent.
LLM Skills for agents
Create Skills your agents can actually use.
Start with a GitHub repo, book, paper, PDF, or video. SkillDiscs turns it into a structured Disk, then gives your agents three ways to use it: portable Skill export, MCP, or REST API.
REST API
Bearer-key auth, JSON in, JSON out. Third-party apps can search, list, and retrieve clean, structured, vector-indexed Disk data.
MCP server
A Model Context Protocol server exposing search_knowledge, get_disk, and list_disks — built for Claude, Claude Code, and other MCP clients.
LLM Skill export
Export any Disk as a downloadable Skill bundle (SKILL.md + references) built from GitHub repos, YouTube videos, papers, books, PDFs, and notes.
Claude
Connect Claude or Claude Code over MCP, or drop in a generated Claude Skill. Answers can cite the exact Disk spans.
Any MCP client
Claude Desktop, Claude Code, Cursor, Windsurf, or your own MCP client — add the endpoint and your Disks become queryable.
ChatGPT & custom GPTs
Point a custom GPT or OpenAI agent at the REST API as a tool/action. It gets the same vector search and structured JSON back.
Your own code
LangChain, LlamaIndex, a cron job, or raw fetch — it is just HTTP + a Bearer key. Deploy anywhere.
API surface
Every production endpoint in one place.
Use REST when you are building a tool or app. Use MCP when your agent client already speaks Model Context Protocol.
POST/api/v1/searchSearch knowledgeSemantic section search across owned Disks, returning relevant hits with key points, source context, and redaction signals.GET/api/v1/disksList DisksDiscover owned and active-saved Disks by scope, category, tag, and metadata.GET/api/v1/disks/{id}Get one DiskRetrieve one Disk by UUID or public_id, including sections, tags, source URL, and conditional source text.POST/api/v1/disks/batchBatch retrieveFetch up to 10 owned Disks in one request for agents that already know the IDs they need.POST/api/mcpMCP serverStreamable HTTP MCP endpoint exposing search_knowledge, get_disk, and list_disks to Claude and other clients.
What you can connect
Four useful ways to give agents SkillDiscs Skills.
The same Disk can become a portable Skill, power a Claude session, back a product feature, or feed a backend workflow.
Claude and Claude Code via MCP
Add the SkillDiscs MCP endpoint and Claude can search your library, open a Disk, and ground answers in your saved sources.
Read MCP setupAny MCP-compatible agent
Use the same tools from Cursor, Windsurf, local agent runtimes, or your own Streamable HTTP MCP client.
View MCP toolsThird-party apps and automations
Call REST endpoints from SaaS apps, workflows, notebooks, or backend jobs that need user-scoped knowledge retrieval.
View REST endpointsLLM Skill packages
Turn one Disk into a portable Skill bundle with instructions, triggers, and references from GitHub, books, papers, PDFs, or videos.
Plan LLM Skill workflowsWhy SkillDiscs
Why agents work more precisely with SkillDiscs
SkillDiscs handles the infrastructure AI needs before it can give good answers.
Human-curated source quality
Every Disk was added because a human found it worth keeping. No spam, no ad-stuffed scrapes, no auto-generated summaries of summaries.
High-level parsing
PDFs, Office files, webpages, YouTube, and GitHub repos become clean sections, tables, figures, headings, and source structure. No 60-page text blob in the context window.
Vector index and retrieval out of the box
Sections and RAG chunks are embedded and stored in pgvector. Your agent queries relevant data semantically without building an embedding or vector database pipeline.
Knowledge base, API, MCP, Skill
The same Disk works as a structured knowledge base, REST API, MCP source, or Claude Skill. Your agent gets data in the right format instead of loose prompt fragments.
Comparison
LLM Skill vs RAG, MCP and a prompt file
Four ways to give a model knowledge. They solve different problems, and they compose.
| LLM Skill | RAG knowledge base | MCP server | Prompt file | |
|---|---|---|---|---|
| What it is | A procedure packaged as files | A corpus, chunked and embedded | A live connection to data and tools | Instructions pasted into the context |
| Where it lives | In the agent's Skills folder | In a vector store | On a server the client connects to | In the conversation |
| How the model gets it | Reads the description, loads the body when a task matches | Retrieved per query by similarity | Called at runtime as a tool | Always present, every turn |
| Best for | Repeatable know-how: a workflow, a house style, a checklist | Answering questions over a large body of text | Data that changes, or actions that must run | One-off or always-on rules |
| Cost of a change | Regenerate the package | Re-embed the changed documents | Deploy the server | Edit and resend |
They stack rather than compete. A Skill can tell an agent to call a SkillDiscs MCP endpoint, and the Disk behind that Skill is the same one the search API retrieves from.
Under the hood
Parsing, vector index, retrieval — ready for agents
SkillDiscs builds the layers agent teams usually have to build themselves: extract, structure, enrich, embed, store, and serve the source.
- 01
Extract
Crawl4AI for URLs, structured parsers for documents/Office/PDF, YouTube transcripts, and GitHub extraction. The output is usable data, not raw files.
- 02
Enrich
Summaries, key points, concepts, knowledge graph, JSON/Markdown structure, and Skill hints make the source readable for humans and models.
- 03
Index
Sections and RAG chunks are embedded and stored in Postgres + pgvector. Your agent searches semantically instead of stuffing context blindly into prompts.
- 04
Serve
REST API, MCP server, and Skill export provide access. Agents can use the Disk as a knowledge base, tool, or Skill.
Live example
See it in action
One POST gets your agent the most relevant sections from across the user's library.
Request
POST https://skilldiscs.com/api/v1/search
Authorization: Bearer sk_live_…
Content-Type: application/json
{
"query": "how does photosynthesis work?",
"limit": 3
}Response
{
"results": [
{
"disk_title": "Plant Biology Fundamentals",
"section_index": 4,
"section_text": "Photosynthesis converts light energy ...",
"key_points": [
"Chloroplasts capture sunlight",
"Calvin cycle fixes CO2"
],
"redacted": false,
"similarity": 0.94
}
]
}Built for
Built for these patterns
Whatever you're building, SkillDiscs replaces raw scrapes and fragile prompt fragments with parsing, vector indexing, retrieval, and structured knowledge bases.
RAG over a curated knowledge base
Your customer support bot, internal copilot, or research assistant gets answers grounded in real, owned content — not generic web data.
Per-user knowledge agents
Each user's agent only sees their library. RLS-enforced. No cross-tenant leakage; perfect for SaaS embedding scenarios.
Drop-in embeddings provider
Skip your own ingest + embedding pipeline. Push raw URLs / PDFs in via the user-facing app, query the API from your backend.
FAQ
Questions about LLM Skills
What is an LLM Skill?
An LLM Skill is a folder an AI agent loads on demand: a SKILL.md file with a name, a description, trigger phrases and a procedure, plus any reference files it points to. The model reads the description to decide whether the Skill is relevant, and only then loads the rest — so the procedure costs nothing until it is needed.
How is a Skill different from a RAG knowledge base?
A Skill carries a procedure and is loaded when its trigger matches the task. RAG retrieves passages by semantic similarity, per query. Put differently: a Skill answers "how do we do this here", RAG answers "what does the source say about this". SkillDiscs produces both from the same Disk, so you do not have to choose upfront.
Can I create an LLM Skill from a GitHub repository?
Yes. Connect the repository and SkillDiscs turns it into a Disk — structure, architecture and a generated wiki — then extracts the procedure from that. PDFs, Office documents, EPUBs, web pages and YouTube videos work the same way.
Which agents can use the Skills SkillDiscs generates?
The package uses the standard Skills folder layout, so Claude and Claude Code load it directly. Any MCP-compatible client can reach the same Disk through the MCP endpoint, and anything that can make an HTTP request can use the REST API at /api/v1 instead.
Does a generated Skill keep a link back to the source?
Yes. Extracted steps keep references to the sections of the Disk they came from, and the Disk keeps its original source attached. An agent using the Skill can point at where a step came from, and query the same Disk through the API for more of the surrounding context.