minouris

Addressing AI Coding Problems Through Constraint-Based Solutions

Generated: 2026-02-24

Summary

The Pitfalls of Vibe Coding document identified four root problems that make AI-assisted coding hazardous: Hallucination (confabulation and fabrication), Dishonesty (false claims of correctness and completion), Overeagerness (premature implementation and unauthorised decision-making), and Amnesia (context loss and forgotten requirements). Each root problem produces operational failures: code that doesn't work, security vulnerabilities introduced without detection, and project abandonment when context poisoning reaches unrecoverable states.

This document examines how constraint-based solutions — instruction files, policy mandates, and compliance gates — were developed to address these problems. The solutions traced here emerged through empirical refinement between October 2025 and February 2026, documenting both failed approaches an

minouris

The Pitfalls of Vibe Coding

What Is Vibe Coding?

Vibe coding is the practice of building software by describing goals in natural language and iteratively prompting an AI model, while relying on minimal review of the generated code. The term was coined by Andrej Karpathy in early 2025. Its defining characteristic is that the developer prioritises speed and feel over understanding: the code works (or appears to work), and that is treated as sufficient.

It is distinct from AI-assisted programming, where the developer understands and verifies each change. By 2025, 25% of Y Combinator startups had codebases written almost entirely by AI tools.

Vibe coding is not inherently wrong. For prototypes, throwaway scripts, and personal tools, the speed advantage is real. The problems arise when the approach is applied to production systems, shared infrastructure, or codebases that must be maintained — or when the person building does not have the engineering background to recognise what is missing.

minouris

AISP Compliance Validation Findings

During validation of Empirical Validation: AISP Policy Formalization Prevents Hallucination, we discovered that the prompts are AISP-inspired but not AISP 5.1 compliant.

Context

The AISP code in the validation document was generated by Claude (Sonnet 4) when explicitly requested to produce AISP-compliant prompts following the AISP 5.1 Platinum specification. While Claude successfully produced parseable and interpretable code that effectively demonstrated policy formalization, it omitted critical specification requirements.

Vanilla AISP Prompt (Non-Compliant)

minouris

Empirical Validation: AISP Policy Formalization Prevents Hallucination

Summary

This document presents empirical findings demonstrating that AISP can effectively reduce LLM hallucination when accuracy policies are formally encoded, while AISP structure alone provides no such protection. These findings validate theoretical claims in the AISP as Metaprompt Compile Target analysis regarding policy formalization.

Table of Contents

• Empirical Validation: AISP Policy Formalization Prevents Hallucination
• Summary

minouris

AISP as a Metaprompt Compile Target: Analysis

Table of Contents

• Executive Summary
• Background
  • What is AISP?
  • Metaprompts in this Repository
• Analysis
• 1. AI Agent Interpretation and Generation of AISP