Giving AI Coding CLIs Full Permission to Run Autonomously - Claude Code, Codex, and Gemini CLI Compared

To let AI coding CLIs run autonomously, you need to configure their permission settings. Claude Code has --dangerously-skip-permissions and auto mode. Codex CLI has --full-auto and --yolo. Gemini CLI has --approval-mode yolo. This post compares the three tools' permission models and explains which mode to use in which situation.

A Tour of Anthropic's 17 Official Skills — What the anthropics/skills Repo Can Do

Anthropic's official Agent Skills repository, anthropics/skills, ships 17 skills as of July 2026. From docx, pdf, pptx, and xlsx office-file manipulation to design skills like algorithmic-art and frontend-design, to developer tooling like mcp-builder and skill-creator, this article summarizes each SKILL.md across five categories.

Should Every Service Ship a Claude Code Plugin? — Rethinking Distribution After codex-plugin-cc

OpenAI's official codex-plugin-cc is genuinely good: I never had to leave Claude Code to drive Codex reviews and investigations. Plugins are working as a distribution channel that delivers a service inside the surface developers already use. Starting from codex-plugin-cc's design, this post discusses which services should ship plugins from day one, and the tradeoffs to weigh before committing.

Adding an `S-N-A` / `S-N-H` Serial Prefix to Epic Sub-Issues to Split Work Between AI Agents and Humans

When we split an Epic issue into sub-issues, each serial number `S-N` represents one work unit. The AI-agent (Claude Code, Codex, and similar) share of that work becomes sub-issue `S-N-A`; the human share becomes `S-N-H`. When one work unit needs both, `S-N-A` and `S-N-H` sit side by side and each closes independently. When one side is enough, we only create that side. We used to draw the split too coarsely inside a single sub-issue: a checklist would mix tasks the agent could handle with tasks only a human could do (rotating production-equivalent credentials, coordinating with an external integration partner, making a UX judgment call). The agent would tick the boxes it could handle, file a PR, and the un-closable sub-issues piled up. Splitting at the sub-issue level while sharing the `S-N` serial keeps the ownership readable on the Epic view. This post covers the actual rules, how we pick the `-A` vs `-H` granularity, and how the prefix carries through to PR titles.

Refactor an existing product fast with Claude Fable 5 and GPT-5.6 Sol first

The previous post's lesson was one I wrote after paying for it. I turned Loop Engineering on for an existing product, and two weeks in, debt really had piled up at ten times the usual speed underneath gates that kept passing. This post is about that regret and the human-driven, high-throughput refactor I went back to run instead, with Claude Fable 5 planning and GPT-5.6 Sol implementing, both on their light settings.

Hand It an Issue, Get Back a PR — A Hermes Agent Skill Looping Between Codex and Claude Code

When agents write the code, the highest-value human work is neither writing nor reviewing it. It is building the foundation the loop runs on, and planning the issue. Shipping this skill made that conclusion concrete.

Pick Your Coding-Agent Model by Opportunity Cost — Why I Run Fable 5 Low and Codex 5.6 Sol Light for Speed

Lately I've deliberately defaulted to light, fast settings: Fable 5 at the Low reasoning level in Claude Code, and 5.6 Sol at the Light reasoning level and the Fast speed level in Codex. Today's high-intelligence models fail with the same shape of mistakes whether you spend a lot of reasoning tokens or a few. If the failure mode is the same, review effort doesn't change, and what matters isn't token cost but wait time, which is opportunity cost. Hard to see, but its dominance is obvious once you accept that humans are the bottleneck.

Reviewing a PR with Claude Code and Codex at Once — Automating Multi-Model Review as a Hermes Agent Skill

A code review from a single model inherits that model's blind spots wholesale. So have independent models read the same diff, and trust the findings where they overlap. Here is how that workflow became a single command.

Health Habits for a Fully-Remote Worker — Closing the 10-Year Fitness Gap with Walks, Mobile AI Coding, and a Sit-Stand Desk

Office workers teleworking four-plus days per week showed a roughly 10-year-equivalent decline in physical fitness compared to non-teleworkers (Meiji Yasuda Health Foundation, December 2025). As someone working fully remote, I want to write down the habits I actually keep: walks scheduled inside work hours, leaning on Claude Code and Codex from mobile, and defaulting to a sit-stand desk.