What does this skill do?

AI Skill Agent-Based Engineering provides a framework for AI agents to perform most software deployment tasks, while humans maintain quality control and risk management. It combines task decomposition into verifiable units, model routing based on complexity, and an evaluation loop (Eval-First) that validates capabilities and regressions before and after each change.

Task Breakdown
Break the work down into 15-minute units, each of which can be verified independently, with a single dominant risk and a clear completion criterion.
Cost-Based Routing
Assign each task to the model that best matches its complexity (Haiku, Sonnet, or Opus) to optimize token cost and execution time.
Evaluation Loop (Eval-First)
Define capacity and regression tests, run a baseline, deploy, and rerun the tests while comparing results to ensure ongoing quality.
Cost and Risk Management
Records the model, token estimate, retries, time, and success/failure status for each task. Scale the model only if the lower-level model fails due to a lack of reasoning.

Usage examples

🧩 Breaking Down a Project
Break down the implementation of a CRUD endpoint into 15-minute tasks, including their completion criteria and risks.
🔀 Route by complexity
I need to add a basic CRUD endpoint and refactor the authentication system. What model do you assign to each task, and why?
🧪 Eval-First in refactoring
Define a performance evaluation and a regression test for the CSV parser, run the baseline, implement the optimization, and compare the results.
📊 Cost Tracking
Record the model, estimated tokens, retries, time, and result for each task in today's session, and let me know where I can cut costs.

Features

Explicit completion criteria Before starting any task, define what it means for the task to be complete to avoid incomplete work.
15-minute intervals Each task can be verified independently, with a single dominant risk and a clear completion criterion.
Model Routing Use Haiku for simple tasks, Sonnet for implementation and refactoring, and Opus for architecture and in-depth analysis.
Eval-First Loop Perform capacity and regression tests before and after deployment, comparing the results to ensure quality.
Cost Audit Generate a log entry for each task, including the model, tokens, retries, time, and success/failure status, to optimize future sessions.

Frequently asked questions

The skill is designed for the Anthropic ecosystem: Haiku for simple tasks (classification, templates, minor edits), Sonnet for implementation and refactoring, and Opus for architecture and in-depth analysis of multiple files.
It is not required. The skill is a methodological framework that is applied within Claude Code or Cowork. If your assessments require running code or tests, you will need access to the corresponding testing system.
The skill logs the model, token count, retries, time, and success/failure status for each task. It only scales up to a higher-level model if the lower-level model fails due to a lack of reasoning, not due to syntax errors.
Yes. You define the completion criteria before each task. The skill guides you to ensure they are clear and verifiable, but it doesn't impose rigid templates.
AI Agent-Based Engineering — Software Development with AI Agents

¿Prefieres escuchar el contenido? Genera la narración de audio con un clic.