skillcrawl

Extract LLM-consumable skills from books, papers, and manuals.

You feed it long-form content. It returns structured SKILL.md files that AI agents can read and execute directly.

How it works

Long-form text  -->  skillcrawl  -->  SKILL.md files
(book, paper,        (extraction      (imperative steps,
 manual, article)     pipeline)        conditions, evidence)

Quick start

Submit content for processing:

curl -X POST https://skillcrawl.sh/process \
  -H "Authorization: Bearer $SKILLCRAWL_API_KEY" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@negotiation-book.epub" \
  -F "granularity=default" \
  -F "type=book"

Response:

{
  "job_id": "sc_7f3a9b2e",
  "status": "processing",
  "estimated_skills": 19,
  "poll": "https://skillcrawl.sh/jobs/sc_7f3a9b2e"
}

Poll for results:

curl https://skillcrawl.sh/jobs/sc_7f3a9b2e \
  -H "Authorization: Bearer $SKILLCRAWL_API_KEY"

{
  "job_id": "sc_7f3a9b2e",
  "status": "complete",
  "skills_extracted": 19,
  "confidence": { "high": 12, "medium": 5, "low": 2 },
  "download": "https://skillcrawl.sh/jobs/sc_7f3a9b2e/skills.tar.gz"
}

Download extracted skills:

curl -O https://skillcrawl.sh/jobs/sc_7f3a9b2e/skills.tar.gz \
  -H "Authorization: Bearer $SKILLCRAWL_API_KEY"

tar xzf skills.tar.gz
ls skills/

tactical-empathy-labeling.skill.md
calibrated-questions-for-discovery.skill.md
mirroring-to-build-rapport.skill.md
accusation-audit-technique.skill.md
...

Output format

Each .skill.md file is an LLM-consumable skill definition:

---
skill_id: negotiation.tactical_empathy.labeling
name: tactical-empathy-labeling
description: >
  Label the other party's emotions to validate them and reduce
  defensiveness. Use when negotiation stalls due to unacknowledged feelings.
type: heuristic
confidence: 0.82
source:
  document: negotiation-book.epub
  chapters: [3, 4]
---

## When to Use
- The other party is visibly frustrated or defensive
- A negotiation has stalled due to unacknowledged feelings

## Steps
1. Observe the emotion without reacting to it.
2. Label it with "It seems like..." or "It sounds like..."
   - Pitfall: never say "I understand" (dismissive)
3. Wait silently after the label.
4. Follow with a calibrated open question.

## Evaluation Criteria
- Success: the other party elaborates after your label
- Failure: they reject the label and escalate

## Evidence
- "Labeling is a way of validating someone's emotion by
  acknowledging it." (Chapter 3, p. 52)

The format is designed for AI agents to load and execute. Human-readable, but machine-first.

Ecosystem

skillcrawl.sh extracts skills from content you own.

skill.broker is the marketplace where extracted skills are bought and sold.

License

Coming soon.