Using Claude Code at Pixery

This guide covers how we use Claude Code at Pixery — best practices, the tools and servers we run, and how to configure CLAUDE.md.

Best Practices

Start long tasks in plan mode. For any non-trivial task, begin in plan mode so Claude proposes an approach before touching code — start with claude --permission-mode plan, or toggle with Shift+Tab / /plan. For a deeply-researched plan you can review in your browser, use /ultraplan <task>.
Default to Opus 4.8 at high effort — it's enough for most work. Reach for /model or /effort only when a task clearly needs more.
Plan with Codex too. When scoping a task, ask the Codex MCP for a second opinion and ideas before committing to an approach.
Separate work, separate sessions. If you have unrelated pieces of work in the same repo, run them in parallel isolated worktrees: claude -w <name> starts a named session; come back with claude -r <name> (or claude -c for the most recent).
Scale agents to the task size: one big task → a subagent; larger multi-part work → agent teams; very large or cross-checked work (audits, big migrations) → ultracode (ultracode in your prompt, or /effort ultracode). See Working with Parallel Agents below for when to use each.
Review with Codex at the end of every task. Always run a Codex code + security review before committing. For a deeper pass, use /code-review ultra (alias /ultrareview).

Working with Parallel Agents

Claude Code can hand work off to other agents instead of doing everything in one conversation. Pick the approach that fits the job (docs):

Approach	What it gives you	Use it when
Subagents	Delegated workers inside one session that do a side task in their own context and return a summary	A side task would flood your main conversation with search results, logs, or file contents you won't reference again
Agent view	One screen to dispatch and monitor sessions running in the background, opened with `claude agents` (research preview)	You have several independent tasks and want to hand them off, check status at a glance, and step in only when one needs you
Agent teams	Multiple coordinated sessions with a shared task list and inter-agent messaging, managed by a lead (experimental, disabled by default)	You want Claude to split a project into pieces, assign them, and keep the workers in sync
Dynamic workflows	A script that runs many subagents and cross-checks their results, for work too big to coordinate one turn at a time (research preview)	A job outgrows a handful of subagents, or you want findings verified against each other: a codebase-wide audit, a 500-file migration, cross-checked research, or a plan drafted from several angles

Important: Dynamic workflows are triggered by putting ultracode in your prompt (or setting /effort ultracode) — use it for the big, cross-checked jobs in the last row.

Claude Code Commands

Handy commands when coding:

Command	What it does
`claude --permission-mode plan`	Start a session in plan mode
`Shift+Tab`	Toggle plan / accept-edits / default mode mid-session
`/plan [task]`	Enter plan mode (optionally with a task)
`/ultraplan <task>`	Draft a deep plan in the cloud, review it in your browser, then execute
`claude -w <name>`	Start an isolated worktree session for parallel work (`--worktree`)
`claude -c`	Resume the most recent session in this directory
`claude -r <name>`	Resume a specific session (no arg → picker)
`/model [opus\|sonnet]`	Switch model / adjust effort (also `--model` at startup)
`/effort <high\|xhigh\|ultracode>`	Set reasoning effort or workflow mode
`ultracode` (prompt keyword)	Run a multi-agent dynamic workflow for one task
`/code-review [--fix]` · `/code-review ultra`	Find bugs & cleanups; `ultra` = deep cloud review (`/ultrareview`)
`/security-review`	Security analysis of the current diff
`/review [PR]`	Review a pull request
`/run`	Launch and drive the app to see a change working
`/verify`	Run the app and confirm a change actually does what it should
`/clear` · `/compact`	Start a fresh conversation · summarize to free up context
`/rewind`	Roll back code/conversation to a checkpoint
`/init`	Generate a CLAUDE.md for the project

MCP Servers

These are the MCP servers we use, ordered by what's required for everyone and what's recommended per role. See Recommended by Role for the full page.

Loading MCP servers...

Plugins

Plugins extend Claude Code with new commands, agents, skills, and integrations. Browse what's available at claude.com/plugins.

Tools like Codex can also be installed as a plugin (instead of, or alongside, an MCP server). Plugins are installed from a marketplace:

/plugin marketplace add <owner/repo>
/plugin install <name>@<marketplace>

If a tool offers a plugin, prefer it over the bare MCP server. A plugin is usually more capable — it can bundle its own slash commands, agents, and skills, not just expose tools — so you get a richer, more integrated experience.

Skills

Skills give Claude reusable, task-specific playbooks. Using them is important — they encode how we work at Pixery, so reach for the relevant skill instead of improvising.

Cloudflare Skills

Install the official Cloudflare skills plugin (important for any Cloudflare work):

/plugin marketplace add cloudflare/skills
/plugin install cloudflare@cloudflare

Pixery Skills

cloudflare-deploy — Deploy projects to Cloudflare Workers with custom domains, D1, R2 (installed by default)

More Pixery skills are added over time; when one exists for your task, use it.

CLAUDE.md

The CLAUDE.md file provides instructions to Claude Code. Project-level ./CLAUDE.md takes precedence over user-level ~/.claude/CLAUDE.md when both exist.

User-Level Template

The installer creates a default user-level CLAUDE.md at ~/.claude/CLAUDE.md. This template is fetched live from the server:

Loading template...

CLAUDE.md Best Practices

User-level for general rules - Keep AI workflow instructions (Context7, Codex reviews, code style) in ~/.claude/CLAUDE.md
Project-level for specifics - Tech stack, project structure, conventions go in project root CLAUDE.md
Don't duplicate - Project CLAUDE.md should only contain project-specific info, not repeat global rules
Keep it concise - Claude reads the entire file, avoid unnecessary content
Be specific - Vague instructions lead to inconsistent behavior
Always review with Codex - After implementing a feature, always use the Codex MCP server to do a code and security review before committing

Updating Configuration

Via Installer

Re-run the installer to update all configurations:

curl -fsSL https://ai-server.pixerylabs.net/install.sh | bash

Manually

Edit the files directly:

# Edit global CLAUDE.md
nano ~/.claude/CLAUDE.md
 
# Edit Claude config
nano ~/.claude.json

Via MCP Tool

Use the Pixery AI MCP server to fetch the latest template:

Get the latest CLAUDE.md template from Pixery