Craft Anything operator
When to use this skill
Use Craft Anything when an answer or rebuild depends on software that was built and shipped:
- Shipped executables and libraries (ELF, PE, Mach-O), WebAssembly, Go symbols, or firmware.
- Packaged desktop apps: Electron apps,
.asararchives, JavaScript bundles, or .NET assemblies. - Mobile packages: Android APKs, DEX code, manifests, and packaged resources.
- Live network packet captures, crash dumps, or sqlite/database files.
- Rebuilding software from an existing website, app, or game using a clean-room workflow where an AI coding agent creates an independent implementation in Rust, Godot, Unity, Unreal, or web stacks.
- Operating the
craft-anything-clicommand-line tool, stdio MCP server, or desktop workbench.
When NOT to use this skill:
- If you already have the source code and the question is answered by reading the source, use normal repository tools (
codebase-search,grep,read). - For generic browser automation on already-logged-in sessions, route to
playwriterorbrowser. - For standard debugging of local project bugs, route to
debugging.
Clean room versus Deep mode
Craft Anything enforces a strict operational distinction:
- Clean room (default): Nothing inside the target program is read — no source, internal bytecode, decompiled functions, or private assets. Craft Anything works exclusively from externally observable facts: screenshots, measured layout, colors, fonts, spacing, and open windows. Tools that inspect internals are refused and return
details.reason: clean_room. During rebuilds, coding agents are kept away from the original: direct reads, directory listings, or execution in the target's folder are off-limits, and the agent builds from harvested captures, references, and user specifications. - Deep mode: Reads internal code, disassembles, decompiles, inspects JavaScript bundles, and connects to Ghidra when installed. Deep mode is for software you own or are explicitly authorized to examine. Activate with
--deepon the CLI orsettings_set deep_mode truein the app.
Instructions
Run the offline gate first. Run the bundled preflight script before attempting upstream execution:
python3 .agent-skills/craft-anything/scripts/preflight.py --jsonAdd
--repo PATH --expect-commit 56aa31976b4ff0950e6b34ba020b0a568d2462fawhen checking a supplied source checkout. Preflight is read-only, makes no network requests, and never executes unverified binaries.Fit-check the installation and tools. Read references/upstream.md and references/cli-mcp.md. Check if
craft-anything-cliis installed:craft-anything-cli --version craft-anything-cli doctor craft-anything-cli capabilitiesIf building from source, verify Rust 1.97+ (
cargo --version,rustc --version). On Debian/Ubuntu desktop builds, ensure required GTK/Wayland system libraries are installed.Rebuild targets deliberately (Study → Build → Check → Ready). Read references/rebuild-workflow.md. To study an existing target and have a coding agent rebuild it:
craft-anything-cli rebuild <target> "<request>" [options]- Target can be a web URL (
https://...), an absolute application/file path, or-for a prompt-only build. - Refinement rounds:
--rounds 2(default). - Coding agents:
--backend "Claude Code"(or Copilot, Grok). - Target engines:
--build-with <engine>(e.g.godot,unity,unreal,rust). - Project location:
--project-folder <dir>or--continue-in <dir>to resume/iterate. - Unattended mode:
--unattendedauto-answers questions with first choice and refuses external approvals.
- Target can be a web URL (
Serve MCP tools safely. To provide Craft Anything's 320 tools to an agent client:
craft-anything-cli mcp --all-tools- Shared host restriction:
--restricteddisables machine-acting tools. - Selectively enable capabilities:
--allow-web,--allow-capture. - Provide an inbox for human approvals:
--inbox <dir>(deposits requests as<id>.request.json, answered via<id>.answer.json). - Setup client configuration: always run
craft-anything-cli setup --dry-runfirst to preview changes. - CRITICAL: Never run
setup --yesover an entry that already carries--all-tools, as plainsetupwrites the core-only command and drops--all-tools.
- Shared host restriction:
Maintain durable investigation memory. When conducting manual or multi-step investigations, persist state across context boundaries:
- Start/resume: call
core_investigation_showandcore_session_show. - Define scope: write brief (
core_brief_write) and initial plan (core_plan_update). - Test hypotheses: record predictions (
core_hypothesis_record), then verify (core_hypothesis_update). - Settle findings: record findings with evidence IDs (
core_finding_record). - Handoff note: write
core_handoff_writeat the end of every turn. - Inspect project artifacts with
python3 .agent-skills/craft-anything/scripts/project_inspect.py inspect <dir>.
- Start/resume: call
Gate consent and side effects.
- Tools that run target code or change devices require explicit user approval (
consent_required). - AI generation of pictures or 3D models with fal.ai (
--generate-assets) incurs API costs and requires explicit budget consent. - Installing dependencies during rebuilds (
npm,cargo) executes external code; always review commands before approving. - For troubleshooting, read references/recovery.md.
- Tools that run target code or change devices require explicit user approval (
Examples
Example 1: Offline preflight check
python3 .agent-skills/craft-anything/scripts/preflight.py --json \
--repo /path/to/craft-anything \
--expect-commit 56aa31976b4ff0950e6b34ba020b0a568d2462fa
Example 2: Clean-room website rebuild
craft-anything-cli rebuild https://example.com "Recreate this landing page in pure HTML and Tailwind CSS" \
--rounds 2 \
--project-folder ~/Documents/Craft\ Anything/example-site
Example 3: Rebuilding a game in Rust
craft-anything-cli rebuild - "A classic arcade snake game in Rust with macroquad" \
--build-with rust \
--rounds 1 \
--project-folder ~/Documents/Craft\ Anything/snake-game
Example 4: Validating an existing project folder
python3 .agent-skills/craft-anything/scripts/project_inspect.py validate \
~/Documents/Craft\ Anything/example-site --json
Best practices
- Distinguish observation, derivation, and inference: Direct bytes and captures are observed; call graphs and CFGs are derived; intent and names are inferred. Label every claim.
- Always cite Evidence IDs (
ev_...): Every finding must link to stored evidence records ininvestigation.cany/. - Clean room by default: Default to clean-room analysis for external software. Only switch to Deep mode when internal inspection is explicitly authorized.
- Never bypass approval gates: Do not automate away
core_action_approveorconsent_requireddialogs. - Respect trademark and licensing limits: ArtCraft brand marks in
docs/brand/are proprietary trademarks; rebuild outputs must not duplicate protected brand names, logos, or commercial assets.
References
- Pinned upstream reference — source commit, architecture, toolsets, and platform notes.
- CLI and MCP server reference — subcommands, toolsets, and MCP flags.
- Rebuild workflow reference — Study → Build → Check → Ready pipeline.
- Connection and recovery reference — troubleshooting and setup recovery.
- Upstream repository
