AI coding guides

Short guides you can follow using the directories on this site.

Termius + tmux: 5 mobile remote-dev tweaks

A rewritten how-to from the TermiusHQ X thread of 26 August 2026. Default tmux is built for a hardware keyboard; on a phone you need mouse mode, a named session that reattaches, a faster Escape, vi copy mode, and a Termius keyboard add-on you actually press. This page quotes the five tips, then expands each one into steps you can do from Termius on iOS or Android.

Complete Claude Code tutorial (2026)

A finishable path from native install through CLAUDE.md, skills, MCP, a Codex comparison, and troubleshooting.

What Claude Code is (and is not)

Name the product correctly, pick one surface, and know what it will not do. Install waits until the next chapter.

Install Claude Code the 2026 way

Use the native installer, confirm PATH, then run version and doctor. Do not start with a global Node package.

Complete OpenAI Codex tutorial (2026)

A finishable course from install to AGENTS.md, skills, MCP, and troubleshooting.

What OpenAI Codex is: ChatGPT, CLI, and IDE

Separate the three doors before you install anything. This chapter names the product and its edges.

How to install OpenAI Codex (CLI, IDE, ChatGPT)

Pick one door, sign in, then run a harmless self-check. Do not configure three clients on day one.

Complete Grok Build tutorial (2026)

A follow-along path from install and plan mode through config, skills, MCP, a Codex comparison, and troubleshooting.

What Grok Build is: TUI, headless, and ACP

Name the three ways of working before you install anything. This chapter is vocabulary and boundaries only.

How to install Grok Build (CLI, login, headless)

Install grok from the official page, finish browser auth or an API key, then do a harmless check. Do not configure three doors the same day.

Install MCP in Cursor

Add one MCP server to Cursor so the agent can call external tools.

Connect MCP in Claude Desktop

Edit claude_desktop_config.json so Claude can use servers from this directory.

Add an Agent skill to Codex

Install one skill so Codex can invoke a fixed workflow by name.

How to pick MCP servers for coding

Choose MCP from real coding needs, not from a popularity list.

How to pick an Agent skill

Filter by category and description. Ignore vanity install counts.

Cursor vs Claude Code vs Codex

Pick a client by editor vs terminal and by the job, not by trending names.

How to write a SKILL.md

A short file that tells the agent when to run, what to do first, and what is forbidden.

MCP security basics: lock permissions first

MCP can read files and run commands. Default trust hands over repo keys.

How to pick offline browser tools

Prefer tools that finish in the tab, especially when pasting sensitive text.

How to use a JSON formatter (and validate)

Expand one-line JSON, minify it, and confirm it parses before you share it.

JWT decode explained: reading expiry

A JWT is three Base64 parts. Decoding reads Header and Payload; it does not verify a signature.

How to use Agent skills in Cursor

Put a skill folder where Cursor can see it, then call the skill by name.

How we curate the AI coding digest

The digest is not a reprint. It is a short list of public items worth opening.

MCP vs Agent skills

MCP exposes tools. Skills expose a method. Most workflows need both.

Offline checklist for sensitive text

Before pasting tokens, logs, or customer data, confirm the work stays on your machine.

MCP vs CLI tools

MCP is a typed tool interface for agents. A CLI is for people and scripts. You often want both.

Cursor Rules vs Agent skills

Rules are always-on constraints. Skills are task-scoped playbooks.

How to debug an MCP connection

Check config, process, permissions, then a tool call. Do not reinstall the client first.

Local vs remote MCP

A local process touches this machine. A remote endpoint touches a shared service. The permission model is different.

Install MCP in VS Code

Add one MCP server to VS Code / Copilot so the agent can call external tools.

Claude Code vs Cursor Composer

One is a long terminal session. The other edits the repo inside the IDE. Pick by where you already spend time.

How to publish an Agent skill

Ship a scannable SKILL.md in a repo people can install. Do not start with a marketing page.

MCP auth and OAuth basics

Local stdio uses env vars on your machine. Remote HTTP endpoints are where OAuth appears. Do not commit tokens.

Toolbox vs CLI for JSON

Use the in-page formatter for a short paste. Use jq for files and pipes. They are not the same job.

Use this site’s offline tools safely

Confirm the work stays in the tab, then paste sensitive text, then clear it.

Install MCP in Windsurf

Add one MCP server to Windsurf so Cascade can call external tools.

Connect MCP in Codex CLI

Add one MCP server to Codex so the terminal agent can call tools.

How to write a minimal MCP server

Expose two or three read-only tools before you talk frameworks, remotes, or auth.

MCP transports: stdio vs SSE vs HTTP

Transport decides who starts the process and where tokens travel. It is not a taste pick.

When you should not install MCP

If one command is enough, do not wire a server. MCP has process, permission, and context cost.

AGENTS.md vs Agent skills

AGENTS.md is repo-level briefing. A skill is an installable, named workflow.

How to evaluate an MCP server before you install it

Check the repo, transport, permissions, and tool count. Ignore vanity star claims.

When MCP tool results blow the context window

Tool output enters the model context. A whole page or a SELECT * will crowd out your code.

How to read an MCP detail page on this site

A detail page is metadata, not an installer. Read transport, author, and the description before you wire anything.

Agent skills vs a prompt library

A prompt is a paragraph. A skill is an installable, named playbook with bans.

Where to put MCP secrets so they stay out of git

Secrets belong in environment variables or the OS keychain. Do not commit sample values from mcp.json.

Which MCP to use when you need docs

Official doc search, a repo README, and a web fetch are different jobs. Ask where the docs live.

Use a skill for your first code review

Install one review skill, give it a diff scope and bans. Do not let it push a branch.

Browser MCP vs fetch

Use fetch for a static public page. Use a browser when you need a login, scripted render, or the console.

How to use the Chinese and English guides together

Each guide has a Chinese URL and an /en/ URL. They are paired pages, not two unrelated machine translations.

How to narrow MCP allowed paths

Filesystem servers take a directory allow list. The whole disk is a key handover.

Pin the MCP package version

npx -y without a version will move under you. Pin what you verified.

Approve MCP tool calls on purpose

A connected server is not a blank cheque. Read the tool name and args before you click allow.

Use MCP for read-only SQL first

A database MCP with a writer role is a production incident. Start with SELECT and a LIMIT.

Install one MCP per task

Ten idle servers spend tokens on schemas you will not call. Keep the set as small as the job.

When you should not install a skill

A one-line reminder belongs in Rules or AGENTS.md. A skill is for a named, reusable playbook.

Share MCP config without sharing secrets

Commit command and args with placeholders. Each person injects env vars locally.

When an MCP server hangs

A hang is usually a process that never answered, not a model that forgot the tool.

How to read a skill detail page on this site

A skill page is metadata: trigger, job, and a project home. It is not an installer that runs in the browser.

CLAUDE.md vs AGENTS.md vs Cursor Rules

Three briefs for the same job: what the agent reads when it enters the repo. Different filenames. Do not write opposite rules.

How to run coding agents in CI

CI has no one to click allow. Stick to read-only checks or a patch on a bot branch. Do not let a headless agent push production.

Set the agent sandbox and approvals

The sandbox limits which files and network the process can touch. Approvals decide when it actually runs. Both should be tighter than always-allow.

Run grok -p headless in scripts

grok -p sends one prompt to stdout so a script can capture it. It is not a substitute for an interactive session.

When claude doctor cannot fix PATH

doctor tells you what is missing. The actual fix is making the GUI terminal and the CI job see the same binary.

Claude desktop vs Claude Code CLI

Desktop is for chat and dropping files. The CLI is for editing a repo in a loop. Pick the window you stare at.

Run agents in parallel with git worktrees

Two agents in one repo mostly fail by stepping on each other's working tree. A worktree gives each task its own directory and branch.

How to review an agent's diff safely

An agent's patch looks tidy, and the risk sits in the lines you skipped. Review it in a fixed order.

Secrets hygiene for coding agents

An agent that can read your repo and run your commands can read your secrets. Assume anything it can read may end up in a log.

When an agent should not touch migrations

A migration is not ordinary code: it runs once in production and is hard to walk back. Most tasks should stop at a draft.

Picking a model for long agent tasks

Long tasks usually fail on context, time and budget, not on raw intelligence. Choose by task shape first, benchmarks second.

MCP or a plain script in CI

CI has nobody to click approve and nobody reading a tool list. Most pipeline steps need one command; MCP pays off only when the model must choose the tool itself.

How to read MCP tool read-only and destructive hints

Annotations are the server's own claims, not a security boundary the client enforces. Read the hints before you approve a tool.

How MCP resources, tools, and prompts split the work

All three can feed outside information into a model, but they differ in how they are called and whether they have side effects. Mixing them turns a read-only catalog into a button that changes data.

How to accept an MCP server on your machine

Leave the editor first and start the process with the same command. If it will not connect, the problem is usually the command, the environment or the transport, not the model.

How to write MCP tool descriptions the model will follow

The model picks a tool from the name, the description and the argument schema. Marketing copy makes it call the tool when it should not.

What belongs in SKILL.md frontmatter

The host reads name and description first, then decides whether to load the rest. Vague frontmatter makes the skill fire on the wrong task.

Using MCP inside GitHub Copilot

Copilot can speak MCP in VS Code, Visual Studio and Copilot CLI. The hosted GitHub server and a local stdio server are not the same config.

Guides FAQ

Do these guides copy official docs?

No. They are original overviews. Project homes remain the source for exact flags and paths.

What should I read after a guide?

Open the MCP or skills directory and practice with an entry that already has install metadata.