Skills and Continual Learning#

What Are Skills?#

Skills are Markdown files that encode successful proof strategies, domain conventions, and common techniques. They are injected into agent prompts before each task, improving results without retraining the model.

Skills live in ~/.eurekaclaw/skills/.

Viewing Your Skills#

eurekaclaw skills

Installing Built-in Skills#

eurekaclaw install-skills
eurekaclaw install-skills --force   # overwrite existing

Installing from ClawHub#

eurekaclaw install-skills steipete/github

Downloads from the ClawHub registry. Requires the clawhub CLI.

How EurekaClaw Learns#

After each session, the ContinualLearningLoop runs automatically. It:

  1. Extracts unique failure patterns from the session

  2. Distills successful proof strategies using the LLM

  3. Writes new .md skill files to ~/.eurekaclaw/skills/

These new skills are automatically used in future sessions.

Learning Modes#

Set EUREKACLAW_MODE in .env:

Mode

What runs after each session

skills_only (default)

Distill failures into new skill files

rl

Skill distillation + Process Reward Model scoring of proof trajectories

madmax

Skill distillation + PRM scoring + cloud LoRA fine-tuning (GRPO)

Writing Skills Manually#

Create a .md file in ~/.eurekaclaw/skills/:

---
name: my_technique
version: "1.0"
tags: [probability, concentration]
agent_roles: [theory]
pipeline_stages: [theory]
description: When to use Azuma-Hoeffding vs Bernstein
source: manual
created_at: 2026-03-21T00:00:00
---

# Azuma vs Bernstein

Use Azuma-Hoeffding when:
- Bounded differences condition holds
- Variance is unknown

Use Bernstein when:
- You have a bound on the variance
- Gives tighter constant factors for small variance

See Skills Reference for the full skill system documentation.