# ACIDBATH

> Production AI engineering patterns for senior engineers and technical leaders.
> Code-first, practitioner-focused content on agentic AI, context engineering,
> and LLM cost optimization.

## What This Site Covers

ACIDBATH provides deep technical content on building production AI systems:

- **Agentic AI Patterns**: Sub-agent architecture, workflow prompts, orchestration
- **Context Engineering**: Token optimization, progressive disclosure, semantic search
- **Claude Code**: Best practices, commands, custom agents, MCP integration
- **Production AI Economics**: Cost analysis, model routing, real-world benchmarks

## Recent Posts

- [My AI Lab on an RTX 3090: What Runs, What Broke, and What Surprised Me](/local-ai-lab-architecture): Full architecture breakdown of a personal GPU inference lab on RTX 3090 — vLLM, SGLang, llama.cpp benchmarks, a Go model gateway with OpenTelemetry, Qwen Scope SAE activations in Grafana, and the FlashInfer investigation that inverted my benchmark results.
- [How 861 Characters Made an AI Agent 41% Better](/personality-eval-harness): Every personality I tested traded one behavioral trait for another. Across 154 runs and five model families, the winners refused the bargain and held the middle.
- [I Was Overloading My System Prompt. An Auto-Researcher Caught It.](/stronger-models-regress-under-heavier-personality): I built an auto-research pipeline in Pi to iterate personality forms across models. The delta table it returned was uncomfortable: the heaviest form lifted one model by 29 points and dragged another down by 15. Trimming the prompt fixed it. Adding to it had broken it.
- [Harnesses: Hard Boundaries for Soft Intelligence](/harnesses-hard-boundaries): A source-level look at how Forge engineers better technical behavior through prompt composition, repo doctrine, task memory, and runtime control loops.
- [865,000 Tokens: Building NomFeed with Opus 4.6](/nomfeed): From 6 billion tokens a month to 865,000 for a complete CLI tool. How architectural discipline around token economics produced 20x efficiency gains.

## Topics

### Infrastructure

- [My AI Lab on an RTX 3090: What Runs, What Broke, and What Surprised Me](/local-ai-lab-architecture): Full architecture breakdown of a personal GPU inference lab on RTX 3090 — vLLM, SGLang, llama.cpp benchmarks, a Go model gateway with OpenTelemetry, Qwen Scope SAE activations in Grafana, and the FlashInfer investigation that inverted my benchmark results.

### Technical Deep Dive

- [How 861 Characters Made an AI Agent 41% Better](/personality-eval-harness): Every personality I tested traded one behavioral trait for another. Across 154 runs and five model families, the winners refused the bargain and held the middle.
- [I Was Overloading My System Prompt. An Auto-Researcher Caught It.](/stronger-models-regress-under-heavier-personality): I built an auto-research pipeline in Pi to iterate personality forms across models. The delta table it returned was uncomfortable: the heaviest form lifted one model by 29 points and dragged another down by 15. Trimming the prompt fixed it. Adding to it had broken it.
- [Harnesses: Hard Boundaries for Soft Intelligence](/harnesses-hard-boundaries): A source-level look at how Forge engineers better technical behavior through prompt composition, repo doctrine, task memory, and runtime control loops.

### Production Patterns

- [865,000 Tokens: Building NomFeed with Opus 4.6](/nomfeed): From 6 billion tokens a month to 865,000 for a complete CLI tool. How architectural discipline around token economics produced 20x efficiency gains.
- [The Three-Layer Pattern for AI in Production Operations](/ai-ops): Why I separated the reasoning engine from the chat interface, and what years of incident response taught me about building AI-augmented workflows.
- [The AI Coding Force Multiplier: Three Patterns That Compound](/force-multiplier): Workflow prompts deliver consistency. Single-file scripts deliver power. Directory watchers deliver scale. Together, they multiply: 1 + 1 + 1 = 10, not 3.

### AI Architecture

- [Where Arabic Morphology Meets AI Engineering: Building a Morphologically-Aware Quran Search System](/qrag): QRAG combines structural Arabic indices with multi-model AI pipelines to transform keyword search into deep linguistic analysis.

### Industry Analysis

- [The Walled Garden Era: What Anthropic's Claude Code Crackdown Means for AI Development](/anthropic-crackdown): Anthropic blocked 650,000 developers at 2 AM without warning. Here's the economics, the enforcement pattern, and what it means for your AI stack.
- [18 AI Engineering Predictions for 2026 (With Confidence Levels)](/predictions-2026): 18 specific AI engineering predictions for 2026, each with confidence levels and falsifiable criteria. Public scorecard dropping January 2027.

### Context Engineering

- [Context Engineering: The Complete Guide for Senior Engineers](/context-engineering): Progressive disclosure, file-based persistence, sub-agent delegation, and Skills vs MCP - reduce tokens by 94%, cut costs 36x.

## About

ACIDBATH is the critical, code-first voice that senior technical leaders trust.
We focus on what actually works in production, including honest coverage of
failures and limitations. No hype, no hand-waving—just working code and
real numbers.

## Contact

- Website: https://amenoacids.com
- Email: contact@amenoacids.com

---

*For full post content, see [/llms-full.txt](/llms-full.txt)*