AaravKashyap12/advise-project-approach

What it solves

This project provides a structured workflow for AI agents to give engineering advice that is grounded in evidence rather than "vibes." It prevents agents from giving generic or fashionable recommendations by forcing them to research actual constraints, comparable real-world projects, and concrete tradeoffs before suggesting a tech stack, architecture, or shipping plan.

How it works

It is implemented as a portable "skill" (defined in a SKILL.md file) that can be loaded into various AI agent harnesses like Claude Code, Codex, and pi. When activated, the skill guides the agent through a specific research loop: understanding the user's goals, inspecting existing codebases, studying credible comparables, and evaluating tradeoffs.

Depending on the project stage, it operates in three modes:

• Pre-build: Researches stacks and architectures to create a build plan.
• Mid-build: Inspects the repository to identify high-impact changes and course corrections.
• Post-build: Reviews the finished project against mature comparables to identify gaps and hardening needs.

Who it’s for

Developers and architects who use AI agents to plan, review, or pivot their software projects and want accountable, evidence-based engineering decisions instead of generic AI suggestions.

Highlights

• Evidence-Based: Requires agents to provide exact dates, adoption signals, and verified sources for claims about project maintenance or popularity.
• Portable: Designed to work across multiple agent harnesses via a standardized skill format.
• Constraint-Aware: Specifically checks for vendor lock-in, operating costs, and the user's specific skill set before recommending tools.
• Guardrails: Explicitly instructs agents to avoid inventing star counts, pretending to review files they haven't seen, or recommending trending tools that don't fit the user's constraints.