Materialize a new project skeleton in one call: .gitignore, .editorconfig, .claude/ (settings, hook, CLAUDE.md, HANDOFF), docs/ by profile (S/M/L), and Auto-memory (work-protocol, product-vision, docs-protocol, MEMORY.md). On M/L profiles also installs the fullstack-architect skill (idea → PRD/ARCHITECTURE/PLANNING/TASKS); S skips it and goes straight to code. The report's next_steps then hands you the remaining skill catalog (frontend-design, humanizer-ru, image, ui-ux-pro-max, emil-design-skills) with a one-line description each — judge from the brief's stack/vision which ones this project actually needs, and offer that pick via AskUserQuestion (mark your recommendations) rather than installing the whole catalog by default. Also checks for the global destructive-command guard (~/.claude, same as install_guard target=user) and installs it only if missing — idempotent, never calls install_guard again once it's there. Idempotent: existing files are never overwritten. Collect the brief (name/stack/profile/vision) in chat first, then call. Note: .claude/ settings/hooks and installed skills load only at session start — if working inside the new project right away, ask the user to restart the session.

Install a personal skill into a project. Bundled skills are copied into .claude/skills// (idempotent, existing files kept); proxied skills are installed by running their official CLI in the project. IMPORTANT: skills load only at session start — after installing, STOP and ask the user to restart the Claude Code app/session before invoking the skill. Note: .claude/skills/ is gitignored by bootstrap_project — skills are tooling, not project code. manual_only applies to bundled skills only.

Set up a website project and hand back a task tracker. One entry point, two flows — pick with kind:

• kind='landing' — build from the bundled generator: an Astro project with a 21-section library (Hero, Problems, Steps, Benefits, Cases, Prices, Reviews, FAQ, Contacts… each with layout variants), a token contract (styles/tokens.css) split from project values (styles/theme.css), a theme picker (/themes) and a section playground (/kit), niche-free page skeletons whose copy is [[…]] markers telling the agent what to write, a privacy page, lead-capture templates (public/send.php + public/assets/lead-form.js), docs/ standards and a machine validator (.claude/check-landing.mjs). Runs npm install. Use for a new site with no reference — one page or several.

• kind='multipage' — rebuild an existing third-party site (donor) as an original project: ships the redesign playbook (donor research, single-source-of-truth data, design tokens, the mandatory redesign delta that keeps the result out of duplicate-content filters, routing, acceptance), battle-tested lead-capture code (send.php + lead-form.js, honeypot, consent, server-side IP, rate limit) plus a tracker seeded with the two research sessions. Deliberately ships NO Astro skeleton and NO template: the donor's structure and stack are unknown until session 1 has looked at it, so the skeleton is raised in session 3. Use when the user points at a site to copy/redo.

Both flows install the same five skills the design work leans on (ui-ux-pro-max, frontend-design, humanizer-ru, image, emil-design-skills), write docs/_dev/tracker.md and append a one-task-per-session protocol to the project's .claude/CLAUDE.md. Skill installs are best-effort — a proxied installer with no network lands in skills_failed while the scaffold still succeeds, so read that field: when it is non-empty, next_steps opens with a fix-it-first step and you install those by hand before starting task 1. Idempotent: existing files are never overwritten; the CLAUDE.md block is appended once.

IMPORTANT — kind is the user's decision. If they only said «сделай сайт» without naming a reference site, ask via AskUserQuestion («лендинг с нуля» / «переделать существующий сайт») before calling.

IMPORTANT — after the call, STOP. Installed skills load only at session start: ask the user to restart the Claude Code app/session, and start work from tracker task 1 in the new session. Do not begin building the site in the same session, and do not run the tracker end to end — one task per session is the point.

Install the destructive-command guard (a PreToolUse/Bash hook) globally into ~/.claude (target=user, default) or into a project's .claude (target=project). Copies destructive-guard.sh into .claude/hooks/ and merges the hook into settings.json WITHOUT clobbering existing keys or hooks (idempotent — safe to re-run). The guard blocks irreversible commands (rm, find -delete, truncate/shred, git reset --hard / clean -f / push --force) while allowing safe variants (git rm, --force-with-lease). Requires jq. Restart the Claude Code session to activate. Note: bootstrap_project already checks for the global (target=user) guard and installs it if missing — you normally only need this tool directly for target='project' or to re-check after a manual jq install.

Generate or edit image(s) via OpenRouter and save them into the project, returning the image inline in chat plus the saved paths and the measured cost. MANDATORY FIRST STEP: the prompt must be written with the bundled 'image' skill — each model needs its own prompt syntax, and Seedream in particular treats comma-separated tags as an anti-pattern. The tool refuses to generate when the skill is missing from the project (it installs it and tells you to read .claude/skills/image/SKILL.md from disk, then call again) or when prompt_source is empty. Default model is Seedream 4.5: $0.04 flat, up to 7.5MP, best prompt adherence — control framing with aspect_ratio and do NOT pass size (that only lowers the resolution for the same price). Seedream covers essentially every task; read the model description before picking anything else, it carries a measured decision table. EDITING: pass the source image via reference_images (local paths or URLs) plus an instruction in the prompt ('remove the sign', 'make the background lighter'); every image model here accepts image input. Name what must stay unchanged explicitly ('keeping its pose unchanged') — that is the vendor-documented way to avoid drift. Mask-based inpainting is NOT supported by this endpoint. Sizing is model-specific: Seedream takes aspect_ratio alone; Gemini needs aspect_ratio + resolution:'2K'. Files land in save_dir (default ./generated, relative to the project). AFTER GENERATING: raw output here is full-resolution and the wrong format for production — run optimize_images on save_dir before shipping (resize/webp/srcset). In a landing build (create_website kind='landing') this is the mandatory last step of the image stage: generate the whole series first (hero → reference_images for the rest, same 'photoshoot'), then one optimize_images call on assets/img at the end — never optimize between individual generations. Requires OPENROUTER_API_KEY in the server .env.

Resize and re-encode images for production web use via sharp: downscale to max_width (default 1920, never upscales), convert to webp (default; jpeg/avif available), auto-apply EXIF rotation, optionally emit extra srcset variants (widths → name-.webp). By default the source file is replaced by the optimized output so heavy originals don't leak into production; keep_originals: true preserves them. Accepts files or directories. Returns final dimensions (use them for width/height) and sizes before/after. Typical landing use: generate images into assets/img, then optimize the whole directory in one call.

Search two local icon sets by English keyword/concept: lucide (~2000 generic UI icons — wrench, shield, clock, printer…) and simple-icons (~3400 brand/company logos — GitHub, HP, Telegram…). Use this before get_icon to find the exact name/set to fetch — exact slugs aren't always guessable.

Fetch one icon's raw SVG markup by exact name from lucide (generic UI icons) or simple-icons (brand logos). Use search_icons first if the exact name/set isn't already known.