askill
gskill

gskillSafety 95Repository

Auto-learn repository-specific skills for coding agents using SWE-smith + GEPA. Generates synthetic tasks and evolves skills through reflective optimization. Use when you want to improve agent performance on a specific repository.

0 stars
1.2k downloads
Updated 3/10/2026

Package Files

Loading files...
SKILL.md

gskill - Auto-Learn Skills for Coding Agents

Automatically learn repository-specific skills by generating synthetic tasks and evolving skills through reflective optimization.

When to Use

  • Agent is making similar mistakes on a repo
  • Want to improve agent pass rate on specific codebase
  • Setting up a new repo for agent work
  • After significant codebase changes

Prerequisites

  • SWE-smith installed at ~/SWE-smith
  • GEPA installed at ~/gepa
  • opencode available in PATH
  • Target repo has tests

Commands

Generate Tasks

gskill generate-tasks --repo ~/prime-radiant-ai --max-tasks 100

Creates ~/prime-radiant-ai/.gskill/tasks.jsonl with synthetic bug-fix tasks.

Evolve Skills

gskill evolve --repo ~/prime-radiant-ai

Runs GEPA optimization loop. Outputs learned skill to .gskill/learned/SKILL.md.

Evaluate Skill

gskill evaluate --repo ~/prime-radiant-ai --skill .gskill/learned/SKILL.md

Tests a skill against tasks to measure pass rate.

Workflow

  1. Generate tasks for your repo (one-time or after major changes)
  2. Evolve skills using GEPA (takes 1-4 hours depending on max_metric_calls)
  3. Review learned skill in .gskill/learned/SKILL.md
  4. Install skill to .claude/skills/learned/SKILL.md in target repo
  5. Measure improvement with gskill evaluate

Output

Learned skills are stored in:

{repo}/.gskill/
├── tasks.jsonl           # Generated tasks
├── evolution_log.jsonl   # GEPA iteration log
└── learned/
    └── SKILL.md          # Learned skill

Example Learned Skill

3) Always check for NULL before aggregation
- Pattern: `COALESCE(column, 0)` or `column IS NOT NULL`
- Failure: NULL propagates, crashes downstream
- Test: `test_null_handling.py`

7) Use idempotent upserts for sync operations
- Pattern: `ON CONFLICT (id) DO UPDATE`
- Failure: Duplicate rows on retry
- Test: `test_sync_idempotency.py`

Configuration

ParameterDefaultDescription
max_metric_calls100GEPA evaluation budget
reflection_modelglm-5LLM for skill reflection
languagepythonTarget language (python, typescript, javascript)

Related Skills

  • skill-creator - For manual skill creation
  • context-* - Repo-specific context skills

References

Install

Download ZIP
Requires askill CLI v1.0+

AI Quality Score

91/100Analyzed 2/24/2026

High-quality technical skill for automated skill learning using SWE-smith and GEPA. Well-structured with clear activation triggers, comprehensive documentation covering commands, workflow, configuration, and examples. Includes proper metadata (tags, description). The skill is reusable across any codebase and provides actionable step-by-step guidance. Minor improvements: could add troubleshooting section and be in a more standard skills folder location.

95
90
88
90
92

Metadata

Licenseunknown
Version-
Updated3/10/2026
Publisherstars-end

Tags

githubgithub-actionsllmtesting