Cost-Aware AI Instructions

Always-on guidance for keeping AI context small, avoiding expensive agent sessions, and classifying tasks before broad edits.

AuthorNeexoCore
Apply to**/*
Updated
costai-creditsguardrails

Overview

These instructions are the baseline for AI-assisted work in Neexo repositories. They help agents keep context small, avoid broad scans, and treat high-risk changes as planning tasks before implementation.

Core Rule

Never combine an expensive model, large repository context, and a vague task. That pattern is the fastest route to unnecessary AI Credit usage.

Task Classes

  • LOW: 1-3 files, clear local change
  • MEDIUM: 3-7 files or uncertain scope
  • HIGH: auth, security, billing, database schema, tenant isolation, deployment, AI/model usage, or broad architecture changes

LOW tasks can be edited directly. MEDIUM tasks should start with a short plan. HIGH tasks should stop after analysis unless explicitly approved.

Tooling Rule

Prefer local CLI tools over MCP servers when a suitable CLI exists. CLIs are often cheaper for context because the agent can ask for --help, inspect only the relevant command, and receive compact output.

Use MCP when the task needs interactive remote capabilities, richer API context, or cross-system access that the CLI does not provide.

If an approved Neexo Windows AI toolbelt CLI is missing and the task benefits from it, install it with winget, verify it with --version, and continue. Ask before installing other tools or modifying PowerShell profile files.

Example: prefer GitHub CLI (gh) for routine issue, PR, repo, and workflow queries. Use GitHub MCP when the work benefits from persistent API tooling or agent-managed GitHub operations.

Raw content

Copy this into your project — e.g. .instructions.md, .agent.md, or SKILL.md

## Overview

These instructions are the baseline for AI-assisted work in Neexo repositories. They help agents keep context small, avoid broad scans, and treat high-risk changes as planning tasks before implementation.

## Core Rule

Never combine an expensive model, large repository context, and a vague task. That pattern is the fastest route to unnecessary AI Credit usage.

## Task Classes

- LOW: 1-3 files, clear local change
- MEDIUM: 3-7 files or uncertain scope
- HIGH: auth, security, billing, database schema, tenant isolation, deployment, AI/model usage, or broad architecture changes

LOW tasks can be edited directly. MEDIUM tasks should start with a short plan. HIGH tasks should stop after analysis unless explicitly approved.

## Tooling Rule

Prefer local CLI tools over MCP servers when a suitable CLI exists. CLIs are often cheaper for context because the agent can ask for `--help`, inspect only the relevant command, and receive compact output.

Use MCP when the task needs interactive remote capabilities, richer API context, or cross-system access that the CLI does not provide.

If an approved Neexo Windows AI toolbelt CLI is missing and the task benefits from it, install it with `winget`, verify it with `--version`, and continue. Ask before installing other tools or modifying PowerShell profile files.

Example: prefer GitHub CLI (`gh`) for routine issue, PR, repo, and workflow queries. Use GitHub MCP when the work benefits from persistent API tooling or agent-managed GitHub operations.