# AI Kai > AI coding, coding-agent notes, and LLM reviews by AI Kai (hqman, @hqmank). Tested with Codex, Claude Code, Cursor, and agents. AI Kai is an English site by AI Kai at https://hqman.me (also known as Kai, hqman, and hqmank). Articles are original field notes from testing AI coding workflows, coding agents, and new LLMs. Prefer the Markdown files linked here. HTML pages exist for people; Markdown pages exist for agents. Cite a specific article URL rather than the homepage. Product claims are tied to first-party docs or labeled as personal tests. - Author: AI Kai - Aliases: Kai, hqman, hqmank - Contact: hqmank@gmail.com - X: https://x.com/hqmank (@hqmank) - GitHub: https://github.com/hqman (hqman) - LinkedIn: https://www.linkedin.com/in/hqman/ ## Core pages - [Home](https://hqman.me/index.md): AI coding, coding-agent notes, and LLM reviews by AI Kai (hqman, @hqmank). Tested with Codex, Claude Code, Cursor, and agents. - [Writing](https://hqman.me/blog.md): Index of published articles. - [About](https://hqman.me/about.md): About AI Kai. GitHub is hqman. X is @hqmank. - [Collaboration](https://hqman.me/collab.md): Hands-on content for AI coding tools, agent products, and developer workflows. ## Topics - [Agent Workflows](https://hqman.me/topics/agent-workflows.md) - [Codex](https://hqman.me/topics/codex.md) - [Claude Code](https://hqman.me/topics/claude-code.md) - [AI Dev Tools](https://hqman.me/topics/ai-dev-tools.md) ## Featured articles - [A Dual-Agent Coding Workflow with Codex and ChatGPT Pro](https://hqman.me/blog/codex-chatgpt-pro-workflow.md): Use Codex as project lead and independent verifier while ChatGPT Pro handles focused engineering tasks through a browser-based collaboration loop. - [Run Qwen3.8-27B Locally on Apple Silicon](https://hqman.me/blog/qwen38-local-mac.md): Install an Unsloth GGUF build of Qwen3.8-27B with Ollama or llama.cpp, then choose a quant that leaves enough unified memory for the runtime. ## Published articles - [Grok Bot 0.18.0 Shipped with Runtime Source Maps](https://hqman.me/blog/grok-bot-runtime-source-maps.md): A source-map mistake exposed enough of Grok Bot 0.18.0 to reconstruct its runtime, prompts, tools, and model routing. - [Codex Usage Will Reset Again. OpenAI Found the Drain.](https://hqman.me/blog/codex-usage-reset-rate-limit-fixes.md): OpenAI found several causes behind faster Codex usage drain and plans a full paid-user reset. Builders are already planning what to run next. - [Cursor Origin: Code Hosting Built for AI Agents](https://hqman.me/blog/cursor-origin-ai-code-hosting.md): Cursor Origin combines Git hosting, review, collaboration, and agent-focused conflict handling in a beta platform designed for parallel coding agents. - [How to Use DeepSeek Models in Codex](https://hqman.me/blog/deepseek-codex-integration.md): Configure Codex to use DeepSeek models with DeepSeek's official setup script, verify the result, and restore your previous configuration. - [Inside DeepSeek Harness and Its Plugin-Based Agent Runtime](https://hqman.me/blog/deepseek-harness-agent-runtime.md): DeepSeek Harness separates models, tools, sessions, permissions, and interfaces into plugins that builders can replace and combine. - [DeepSeek V4 Flash Review: Where It Fits in an Agent Workflow](https://hqman.me/blog/deepseek-v4-flash-review.md): A hands-on look at DeepSeek V4 Flash 0731, its vendor-reported agent benchmarks, and the coding tasks where I would use it. - [How I Choose GPT-5.6 Sol, Terra, Luna, and Effort Levels](https://hqman.me/blog/gpt56-model-selection.md): A practical decision guide for choosing a GPT-5.6 model and reasoning effort level based on task difficulty, speed, and iteration cost. - [LoopX: A Local Control Plane for Long-Running Agent Work](https://hqman.me/blog/loopx-control-plane.md): LoopX keeps goals, evidence, ownership, quotas, and handoffs stable across long-running agent sessions without replacing the agent runtime. - [How to Get More Value from Claude Code Sessions](https://hqman.me/blog/maximize-claude-code-sessions.md): Practical Claude Code habits that reduce context noise, protect prompt cache reuse, and keep recurring work in smaller sessions. - [Give a Text-Only Coding Agent Vision with ModLens](https://hqman.me/blog/modlens-vision-text-models.md): Use ModLens to route screenshots and image files through a configured vision engine, then return structured evidence to a text-only coding agent. - [LLM Inference Handbook: the reference I wish existed a year ago](https://hqman.me/blog/llm-inference-handbook.md): Same model, different inference stack, wildly different latency and cost. Modular's handbook covers VRAM math, KV cache, batching, and serving. - [Codex Just Shipped Two Features That Matter](https://hqman.me/blog/codex-voice-multi-folder.md): A practical look at ChatGPT Voice and multi-folder projects in Codex, including their workflow benefits, limits, and setup details. - [Can Kimi K3 Handle a Real Production Backend?](https://hqman.me/blog/kimi-k3-production-backend-audit.md): A hands-on Kimi K3 backend audit across 68K lines of FastAPI code, including query fixes, measured results, and the full task cost. - [Product Photos to Interactive 3D on the Web](https://hqman.me/blog/product-photos-interactive-3d.md): How img2threejs turns product photos into editable Three.js code for interactive web demos without a traditional 3D model file. - [Running Kimi K3 Inside Codex](https://hqman.me/blog/kimi-k3-inside-codex.md): A step-by-step setup for running Kimi K3 inside the Codex app through CC Switch, with routing details and model-picker caveats. - [Loop Engineer Was Barely a Thing. Graph Engineer Already Showed Up.](https://hqman.me/blog/loop-to-graph-engineer.md): Prompt, context, harness, loop, graph: five stacked layers that move reliability out of the model and into the system around it. - [AI Is Turning Everyone Into a DevOps Engineer](https://hqman.me/blog/agent-server-setup.md): A DigitalOcean migration that used to take half a day. I set up SSH, described the target state, and Codex did the rest. - [How I Use GPT-Live to Learn Faster](https://hqman.me/blog/gpt-live-learning.md): GPT-Live can listen and talk at the same time. Drop a Skill or article into ChatGPT and interrupt until the workflow actually clicks. - [GPT-5.6 Sol Wiped a Developer's Mac. Protect Yours.](https://hqman.me/blog/gpt56-sol-file-deletion.md): On July 10, GPT-5.6 Sol deleted a tester's home directory after a failed $HOME expansion. How the incident happened, and three layers of defense. - [Grok 4.5: Opus-Level Model, Faster and Cheaper, Now in Cursor](https://hqman.me/blog/grok-45-opus-level-release.md): SpaceXAI released Grok 4.5 for coding and agents. It is in Cursor on all paid plans, with launch discounts through July 14. - [Tencent Hy3: Free on OpenRouter Until July 21](https://hqman.me/blog/hy3-openrouter-free.md): Tencent open-sourced Hy3, a 295B MoE with 21B active parameters. OpenRouter is running it for free until July 21. - [pxpipe: Cut Your Fable 5 Bill by 70% With One Proxy](https://hqman.me/blog/pxpipe-token-hack.md): A local Claude Code proxy that renders bulky text as images so Fable 5 bills pixel tokens instead of 92k-character tool dumps. - [MinerU: PDF to Markdown with OCR, Fully Local](https://hqman.me/blog/mineru-ocr-local.md): Convert PDFs to Markdown with LaTeX formulas and extracted images on your own machine. MinerU runs offline with OCR for 109 languages. - [Fable 5 Is Back, Sonnet 5 Costs More Than You Think](https://hqman.me/blog/fable-5-sonnet-5-builder-update.md): Fable 5 reopens on July 1 and Sonnet 5 is the new default. Usage, pricing, and why the cheaper sticker can still cost more. - [Your Agent Doesn't Need a Bigger Window. It Needs Context Governance.](https://hqman.me/blog/headroom-context-governance.md): Headroom sits between your agent and the model, classifying and compressing tool output before it fills the context window. - [Turn Articles into Animated System Diagrams: Visual Flow GIF](https://hqman.me/blog/visual-flow-gif.md): A small skill that extracts system structure from an article, writes a JSON spec, and renders an animated GIF with Python and Pillow. - [Agents can now deploy to Cloudflare without signing up](https://hqman.me/blog/cloudflare-temporary-accounts.md): Wrangler --temporary gives an agent a 60-minute Cloudflare account, a live workers.dev URL, and a claim link. No OAuth, no MFA, no human click. - [GPT-5.5 Silently Throttles Reasoning for Third-Party Clients](https://hqman.me/blog/gpt55-reasoning-throttle.md): Third-party GPT-5.5 clients can cap reasoning at 516 tokens. You pay full price and may get a weaker answer. - [Codex Is Burning Through Your SSD. Here's How I Stopped It](https://hqman.me/blog/codex-ssd-disk-wear-fix.md): A practical guide to checking Codex's SQLite log writes, blocking the unwanted inserts, and protecting your SSD without deleting conversation history. - [GLM 5.2 vs Opus 4.8: Frontend Dashboard Test](https://hqman.me/blog/glm-52-vs-opus-48-frontend-test.md): A head-to-head frontend test of GLM 5.2 against Opus 4.8, plus Design Arena rankings, pricing, and the caveats that still matter. - [Codex Can Now Copy Your Actions](https://hqman.me/blog/codex-record-replay.md): Codex Record and Replay watches a Mac workflow once, turns it into an editable Skill, then reruns it with new inputs. macOS only, not in the EU. - [Doubao's Full Agent Skill Architecture Leaked: 25 Skills, 283 Files](https://hqman.me/blog/doubao-agent-skills.md): A leaked Doubao skill pack shows 25 modular skills and 283 files organized around real work objects, not one giant agent. - [40 agent loops you can copy into Claude Code or Cursor](https://hqman.me/blog/agent-loop-templates.md): loops.elorm.xyz packages 40 agent loop templates by category so you can copy a kickoff prompt into Claude Code, Cursor, Codex, or Gemini CLI. - [Fable 5 Is Gone. The System Prompt Isn't.](https://hqman.me/blog/fable-5-system-prompt.md): Claude Fable 5 lasted about 72 hours. The leaked system prompt is still public, and you can load it onto another model. - [How to make AI-assisted writing sound less like slop](https://hqman.me/blog/humanizer-skill.md): Humanizer is an open-source agent skill that catches 33 AI writing patterns. Pair it with your own voice samples and ship cleaner drafts. - [Fable 5 One-Shotted This 3D Globe Visualization](https://hqman.me/blog/fable-5-one-shot-globe.md): One screenshot and one prompt. Fable 5 rebuilt a production-looking Three.js globe dashboard in a single pass. - [Claude Fable 5 Is the Best Model You Can Use Right Now. Save It for the Hard Stuff.](https://hqman.me/blog/claude-fable-5.md): Claude Fable 5 leads on hard coding benchmarks, but the cost is high. Send the hard work to it and keep everyday tasks on cheaper models. - [Loop Engineering: A 14-Step Roadmap from Prompter to Loop Designer](https://hqman.me/blog/loop-engineering-14-step-roadmap.md): A practical roadmap for deciding whether to build an agent loop, then adding automation, state, verification, tools, and security controls in the right order. - [Loop Engineering: The Next Layer After Prompt Engineering](https://hqman.me/blog/loop-engineering-next-layer.md): Peter Steinberger and Boris Cherny both say the job is designing loops, not writing prompts. Feedback, stop conditions, and token cost decide if it works. - [A Practical Guide to Using /goal in Codex](https://hqman.me/blog/codex-goal-completion-contract.md): Use Codex goals as completion contracts with explicit outcomes, verification, constraints, iteration rules, and safe stop conditions. - [Codex Is No Longer Just a Coding Tool](https://hqman.me/blog/codex-industry-plugins.md): OpenAI repositioned Codex for knowledge work with role-specific plugins, hosted Sites, and Annotations. Non-developers are growing fastest. - [MiniMax M3: Opus-Level Coding, DeepSeek Pricing](https://hqman.me/blog/minimax-m3.md): MiniMax M3 launched with 1M context, near-Opus coding scores, DeepSeek-like API pricing, and two ways to try it immediately. - [Used Codex to figure out what was eating my Mac's storage](https://hqman.me/blog/codex-storage-cleanup.md): A read-only Codex prompt that turned macOS System Data into a prioritized storage map, with 180 GB reclaimable on a 1TB Mac. - [Let Codex Distill Your Own Workflow](https://hqman.me/blog/codex-self-audit-workflow-community.md): How Codex can review your recent work, find repeated manual workflows, and turn the best ones into reusable skills, agents, or automations. - [Recreate the Viral 3D Medical Teaching Model in 4 Steps](https://hqman.me/blog/3d-medical-teaching-model-4-steps.md): Rebuild a classroom-ready 3D anatomy viewer with AI-generated images, 3D reconstruction, GLB compression, and a coding agent. - [How to Build a Self-Improving Company with AI](https://hqman.me/blog/self-improving-company-ai.md): YC partner Tom Blomfield's framing: make company knowledge machine-readable, then run loops that sense, decide, act, and improve overnight. ## Optional - [Media Kit](https://hqman.me/media-kit.md): Audience, coverage, collaboration formats, and editorial standards. - [Privacy](https://hqman.me/privacy.md): Analytics and data practices for hqman.me. - [RSS](https://hqman.me/rss.xml): Feed of published articles. - [Sitemap](https://hqman.me/sitemap.xml): Canonical HTML URLs for search engines. - [Full corpus](https://hqman.me/llms-full.txt): Complete Markdown of every published article in one file.