dot_codex

Imported from ybarsotti/dotfiles (dot_codex/AGENTS.md). Install upstream with npx skills add ybarsotti/dotfiles --skill dot_codex. Copyright stays with the author.

Global AGENTS.md

@RTK.md

graphify

• graphify (~/.Codex/skills/graphify/SKILL.md) - any input to knowledge graph. Trigger: /graphify
• When user types /graphify, invoke Skill tool with skill: "graphify" before anything else.

Browser and Web Navigation

Three browser tools serve three targets. Pick by where the page lives.

• kitesurf MCP — the public web. Cloudflare's remote headless browser for agents. It runs on Cloudflare Workers, so this machine installs no browser and the request leaves from Cloudflare. Use it to open a public URL, read a page, fill a form, click an element, take a screenshot, or scrape. It reads CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN from the environment. If it fails to authenticate, report the missing variable and stop. Docs: https://developers.cloudflare.com/browser-run/kitesurf/
• obscura MCP — this machine's network. A local headless browser written in Rust, with no Chrome. Use it for a localhost dev server, a private address, or any page kitesurf cannot reach from Cloudflare. obscura serve --port 9222 also exposes a CDP endpoint that Playwright and Puppeteer connect to. Docs: https://github.com/h4ckf0r0day/obscura
• A real browser — for the user's logged-in profile, an Electron application, or QA evidence such as screenshots and video.

Prefer kitesurf for a public page. It costs this machine nothing and starts fastest.

caveman

• Activate caveman skill in full mode by default every session.
• /caveman lite|full|ultra overrides current level.
• stop caveman or normal mode disables it for current session.

Writing Style — Simplified Technical English

Anything you write that outlives the conversation follows ASD-STE100 (Simplified Technical English): docs, code comments, commit messages, PR and issue text, plans, reports, and memory files. Write complete sentences with articles. This prose is not chat.

Rules

• One instruction per sentence. Procedural sentences hold 20 words or fewer, descriptive ones 25.
• Paragraphs hold 6 sentences or fewer, and cover one topic each.
• Use the active voice and the present tense. Write "the script reads the manifest", not "the manifest is read".
• One word carries one meaning. Choose a term and reuse it. A lane stays a lane; never rename it for variety.
• Keep noun clusters to 3 words. Write "the script that validates run state", not "the run state validation script".
• State technical facts plainly. Do not use idiom, metaphor, or slang for them.
• Write sequential actions as separate sentences or numbered steps. Do not chain them into one sentence.
• Say what to do. Prefer the positive instruction when it carries the same rule as the negative one.

Length

• Give the answer first. Add the support after it. Do not write a preamble and do not restate the question.
• Do not summarise work that the diff or the tool output already shows.
• Delete a sentence that only signals effort, such as "I carefully reviewed every file".

Precedence

• Chat replies: caveman wins while it is active. Fragments and dropped articles are its purpose. STE governs the persisted artifacts that caveman already exempts.
• Code: ponytail decides what gets built. STE shapes only the prose around it.
• Never compress away a negation, a number, a unit, an exact error string, or a safety warning. Brevity never outranks accuracy.