skills.rest
PrimeIntellect-ai
PrimeIntellect-aiOFFICIAL
Explore all skills by this author →
Website𝕏@primeintellectGitHub

synthesize-task

Generate tiered general-agent task families with empirical pass-rate gating.

113|35|Updated Nov 29, 2025
Search Tags:#database schema#tool orchestration#agent evaluation#semantic verification#task synthesis#difficulty curriculum
One-click install
npx skills add https://github.com/PrimeIntellect-ai/research-environments --skill synthesize-task
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: synthesize-task
Source: https://github.com/PrimeIntellect-ai/research-environments/tree/main/environments/general_agent/skills/synthesize-task
Command: npx skills add https://github.com/PrimeIntellect-ai/research-environments --skill synthesize-task

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Synthesize a new general-agent task family that reliably grows in difficulty across tiers while maintaining measurable, empirical pass-rate control.

Core Features & Use Cases

  • Task Family Generation: Build a tiered sequence of tasks (tier 0 → tier N) where each higher tier is a strict superset of the previous one.
  • Semantic Verification Requirements: Enforce a required verify(db) -> float function so success is judged by semantic correctness rather than matching a single gold path.
  • Difficulty Gating with Pass-Rate Measurement: Use empirical rollouts (exactly 20 for tier seeding, 20 per tier) to ensure each tier lands in the intended difficulty band.

Quick Start

Run the general-agent synthesis workflow described in this skill to create a new task family under environments/general_agent/tasks/ and iteratively validate its tier difficulty using the required evaluation gates.

Frequently Asked Questions about synthesize-task

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I generate tiered agent tasks with empirical pass-rate gating?▼

Generate tiered agent tasks by transforming a seed task into progressively harder instances with strict superset constraints, using 20 empirical rollouts per tier to verify pass-rate difficulty bands.

How does semantic verification work for general-agent task evaluation?▼

Semantic verification uses a required `verify(db) -> float` function to judge task success by semantic correctness rather than matching a single gold tool-call path.

What files are needed to create a general-agent task family?▼

Creating a task family requires a root `SKILL.md`, tiered task directories containing `task.toml`, `instruction.md`, `db.json`, `tools.py` with a `verify(db)` function, and gold tool-call chains.

Can I use existing database schemas and tool definitions to build escalating difficulty curricula?▼

Yes, you can expand a general-agent task set by transforming a seed task into stronger database schema instances, tool usage constraints, and verification logic across escalating difficulty tiers.

Why does my task synthesis difficulty gating require exactly 20 rollouts per tier?▼

Difficulty gating uses exactly 20 empirical rollouts for tier seeding and 20 per subsequent tier to reliably measure pass-rates and ensure each tier lands in the intended difficulty band.

What is the best way to validate general-agent task difficulty after synthesis?▼

Validate synthesized task difficulty by running the `general-agent validate` and `vf-eval` difficulty checks against the generated tiered task directories and gold tool-call chains.

Related Skills

Explore More Education & Research Skills→
MomoDavilukeMomoDaviluke
2

Database Designer

Generate normalized database schemas, migrations, seed data, and ER diagrams.

Community
Intermediate
samymitysamymity
1

eval-harness

Implement eval-driven development workflows for Claude Code with pass@k tracking.

Community
Intermediate
marcosd4hmarcosd4h
17

import-export-resolver

Resolve PE import and export relationships across analyzed modules.

Community
Advanced
junhyekhjunhyekh
1

rlxp-curriculum-design

Design curriculum adjustments for reinforcement learning tasks using per-bin evidence.

Community
Advanced