PLATFORM GUIDE · CLAUDE CODE / COWORK

Run a Loadout
in a workspace

Claude Code and Cowork suit Loadouts that read files, create artifacts or call tools. Review the project boundary and approval gate before allowing writes or external actions.

Connected flow · verify permissionsAILoadout v0.2 draft

Follow the four steps Compare all AI clients

RUN IT

Four controlled steps

The interface differs; the safety contract stays the same.

Open the correct folder

Keep the workspace scoped to the intended project, never a broad parent directory.

Have the AI read the file

Reference the `.loadout.md` path or paste it, then ask for Purpose, Risk and dependencies before execution.

Inspect MCP and write access

Install only reviewed sources and keep filesystem and account scope as narrow as possible.

Approve in stages

Review diffs, exact payloads and test evidence before approving writes, sends or destructive actions.

START COMMAND

Tell the client what must not drift

Paste this after attaching or pasting the Loadout.

Load this Loadout and follow it exactly.
1. Summarize Purpose, Risk, prerequisites and unavailable tools.
2. Ask only for missing inputs before doing the Steps.
3. Never skip Never, approval or Done means.
4. Use the declared output format; fall back to structured Markdown.

CAPABILITY

What is actually supported

Status separates the portable core from account-dependent capabilities.

Core Loadout

Reads Markdown in the workspace and follows ordered work

File-compatible

Project files

Can work against files and inspect diffs

Strong fit

MCP

Servers must be configured and scoped before use

Installable support

Visual

Markdown, Mermaid, HTML or Artifact when available; never color alone

Surface-dependent

LIMITS

Know when to stop

A correct blocked result is safer than pretending a missing capability exists.

Claude web, Claude Code and Cowork are not one identical runtime.
A file write may be reversible; push or publish is an external action.
Cowork custom visuals have sharing and interaction limits.

PRIMARY SOURCES

Product facts, not assumptions

Capabilities can change; check these sources and the tested date.

AnthropicMCP in Anthropic products

AnthropicCustom visuals in chat and Cowork

AnthropicArtifacts

Core flow tested by AILoadout on 23 Jul 2026; connected capabilities remain account- and surface-dependent.