What is an eval-first loop?

An eval-first loop requires defining capability and regression evaluations before execution, running a baseline to capture failures, and then verifying the implementation against those metrics.

What is the 15-minute unit rule in agentic engineering?

The 15-minute unit rule is a decomposition strategy where every task is broken down into a small, independently verifiable unit with a single dominant risk and a clear 'done' condition.

How does this skill handle different Claude models?

It uses cost-aware routing: Haiku for classification and narrow edits, Sonnet for core implementation, and Opus for complex architecture and root-cause analysis.

When should I use the compact vs. fresh session strategy?

You should continue sessions for closely-coupled units, start fresh sessions after major phase transitions, and only compact context after milestone completion to avoid losing active debugging data.

Agentic Engineering Workflow

Name: Agentic Engineering Workflow
Author: shihyu

byshihyu

•

Productividad y Flujo de Trabajo

Operates as a high-efficiency agentic engineer using eval-first execution, task decomposition, and cost-aware model routing.

This skill transforms Claude into a sophisticated agentic engineer designed to handle complex implementation work with minimal human friction. It enforces a rigorous 'eval-first' methodology where success criteria and regression checks are defined before any code is written. By decomposing work into 15-minute verifiable units and intelligently routing tasks between Haiku, Sonnet, and Opus models, it optimizes for both technical precision and cost-efficiency. It is particularly useful for large-scale refactors, architectural changes, and projects requiring high-integrity code reviews focusing on invariants and security boundaries.

Características Principales

01Systematic 15-minute unit task decomposition strategy

025 GitHub stars

03Optimized session and context management for long-running workflows

04Eval-first execution loop for measurable progress and regression safety

05Security-focused code review priorities for AI-generated logic

06Cost-aware model routing based on task complexity (Haiku/Sonnet/Opus)

Casos de Uso

01Optimizing LLM token usage and costs during large-scale development sessions

02Managing complex multi-file refactors with strict invariant requirements

03Automating high-volume feature implementation using test-driven development

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add shihyu/jason_note agentic-engineering

For use in Claude.ai and ChatGPT

Características Principales

01Systematic 15-minute unit task decomposition strategy

025 GitHub stars

03Optimized session and context management for long-running workflows

04Eval-first execution loop for measurable progress and regression safety

05Security-focused code review priorities for AI-generated logic

06Cost-aware model routing based on task complexity (Haiku/Sonnet/Opus)

Casos de Uso

01Optimizing LLM token usage and costs during large-scale development sessions

02Managing complex multi-file refactors with strict invariant requirements

03Automating high-volume feature implementation using test-driven development