Migrate from claude to claude-code.
10 hands-on-verified translation patterns — what carries over and what to watch for. Cited to the Feature Parity Map; the audit tells you whether the move is worth it.
Skills follow the Agent Skills open standard (agentskills.io) and work in both products with the same SKILL.md shape: YAML frontmatter (`name`, `description`, `when_to_use`, etc.) plus Markdown instructions and optional scripts. On Claude.ai you install a skill by uploading the zipped folder under Settings → Capabilities → Skills (Team/Enterprise can publish to an org directory). In Claude Code you place the skill folder at `~/.claude/skills/<name>/` (personal) or `.claude/skills/<name>/` (project, checked in) and invoke with `/<name>` or via auto-trigger. Same skill artifact, different distribution: Claude.ai uses upload-to-account, Claude Code uses files-on-disk.
- Warning: Claude Code-only frontmatter keys (`allowed-tools`, `disallowed-tools`, `context: fork`, `agent`, `hooks`, `shell`, etc.) may be ignored or behave differently on claude.ai — the more agentic surface area on Claude Code lets skills do more.
- Warning: Bundled Claude Code skills (`/code-review`, `/debug`, `/loop`, `/batch`, `/claude-api`, `/run`, `/verify`) are not auto-present on claude.ai; the reverse is also true for chat-flavored skills.
- Warning: On Team/Enterprise Claude.ai, org admins publish skills to a directory; in Claude Code, project skills are shared via git and enterprise skills via managed settings — set the right scope.
- Warning: Claude.ai surfaces skills inside chat and Cowork; Claude Code surfaces them inside CLI/IDE sessions — the same skill folder usually works in both, but always test.
Claude.ai Artifacts are self-contained pieces of code (HTML, React, SVG, markdown, single-file apps) that render in a side panel next to the chat, are iteratively editable, and can be shared/published as live apps. Claude Code has no Artifact concept: when the agent produces code, it writes real files to disk via the Write tool, and you review the diffs (terminal or VS Code diff viewer). To 'migrate' an artifact workflow, ask Claude Code to scaffold the same code as a file in your project, then preview it via your local dev server or by opening the HTML file.
- Warning: No live preview pane in Claude Code — you preview by running the file locally (e.g., `python -m http.server`, `npm run dev`) or by opening the HTML in a browser.
- Warning: Artifact 'Publish' (creating a shareable claude.site URL) has no Claude Code equivalent; share by committing to a repo and deploying yourself (Vercel, GitHub Pages, etc.).
- Warning: Artifacts on Claude.ai have access to a curated runtime (React, Tailwind, a few libraries); Claude Code writes to your real project where you control the toolchain — the same prompt may produce different code shapes.
- Warning: Iterating on an artifact happens in-chat via diffs; in Claude Code the iteration loop is Edit tool → diff review → re-run — slower for tiny visual tweaks, much stronger for multi-file changes.
Claude.ai automatically saves every chat to the sidebar, supports chat search across history, and offers an opt-in Memory feature that synthesizes recurring context across chats (paid plans get richer recall; free has basic memory; Project memory is scoped to the Project). Claude Code is session-based: each `claude` invocation starts fresh and reads the project's CLAUDE.md on every turn, but there is no implicit cross-session memory of past conversations. To carry context forward, either resume the exact session (`claude --resume`) or codify durable facts into CLAUDE.md / sub-agent MEMORY.md via the `#` quick-add shortcut.
- Warning: Claude.ai 'Memory' is automatic and continuously updates; Claude Code 'memory' is a static file you (or the agent at your direction) edit — they are not the same mechanism.
- Warning: Cross-machine: claude.ai history syncs across devices automatically; Claude Code sessions live under `~/.claude/` on the local machine — back up or check in a dotfiles repo if you need portability.
- Warning: Sub-agent MEMORY.md (Claude Code) gives a single sub-agent persistent notes across conversations; this is much narrower than Claude.ai's broad chat-history search.
- Warning: Claude.ai Incognito chats skip memory entirely; Claude Code has no incognito toggle — to keep a session out of `~/.claude/projects/...` history, you would need to use a throwaway working directory or clear history manually.
On claude.ai you click the attach button (or drop files into the chat) to upload PDFs, images, code, or office docs — files are stored in Anthropic's cloud and made available to the conversation (or to a Project's knowledge base). Claude Code does not upload to the cloud; instead it reads files directly off your local filesystem via Read/Glob/Grep, and you point it at files with `@path/to/file` or by giving it a directory. The migration is conceptual: move from 'upload this PDF to the chat' to 'put the file in a known directory and let the agent Read it'.
- Warning: Image and PDF parsing in Claude Code uses the Read tool's built-in PDF/image support (per-page reads on large PDFs), not a vision endpoint upload — large PDFs require the `pages` parameter.
- Warning: Claude.ai per-chat file caps and Project-knowledge size limits do not apply in Claude Code; instead the constraint is the model's context window plus your token budget.
- Warning: Sensitive files: claude.ai uploads go through Anthropic's infrastructure (and may live in a Project); Claude Code only sees files inside the working directory plus anything you explicitly Read — review managed-settings/permissions if you don't want certain paths accessed.
- Warning: Claude.ai supports drag-and-drop in any chat; in Claude Code you have to either place files in the project or paste paths/URLs.
Both products speak the same Model Context Protocol — only the management UX differs. On claude.ai you enable/configure connectors from Settings → Connectors (Slack, Google Drive, GitHub, Linear, Asana, etc.) with OAuth flows in the browser. In Claude Code you add servers with `claude mcp add` (stdio/HTTP/SSE transport) or edit the MCP config file, then manage in-session with `/mcp`. Many of the same servers (GitHub, Slack, Drive) work in both; for instance `claude mcp add --transport http github https://api.githubcopilot.com/mcp/ --header 'Authorization: Bearer ...'` mirrors a Claude.ai GitHub connector.
- Warning: Claude.ai connectors run as managed cloud connections — OAuth happens once and is bound to your Anthropic account; Claude Code MCP servers run locally (stdio) or as remote HTTP/SSE, and you manage credentials yourself.
- Warning: Interactive 'MCP Apps' UIs added on Claude.ai in 2026 render rich connector UIs inside chat; Claude Code surfaces MCP tools as plain tool calls (no embedded UI) — the same server may look very different across the two.
- Warning: Per-skill/per-subagent allowlists (`mcpServers` frontmatter) only exist in Claude Code — Claude.ai gates connectors at the account/org level.
- Warning: The built-in `ide` MCP server in Claude Code (diagnostics + Jupyter cell execution) has no claude.ai analog.
Both products expose Anthropic's frontier lineup (as of May 2026: Opus 4.7, Sonnet 4.6, Haiku 4.5) with the same naming. On claude.ai you choose the model from the chat composer dropdown. In Claude Code, use `/model` to list and switch mid-session, optionally with `effort` levels (`low`/`medium`/`high`/`xhigh`/`max`); the `opusplan` mode plans with Opus and executes with Sonnet. Per-skill or per-sub-agent `model:` frontmatter overrides apply for the duration of that turn.
- Warning: Same model names but Claude Code adds per-tool / per-skill model overrides; claude.ai is conversation-scoped only.
- Warning: Opus burns through plan quota much faster than Sonnet on both products — and the quota pool is shared between chat and Claude Code.
- Warning: Team/Enterprise admins can restrict which models appear in the picker; that restriction propagates to Claude Code when you log in with the org account.
- Warning: Haiku is great for mechanical Claude Code work (renames, formatting batches) but typically not surfaced as a default option on Claude.ai chat — pick it explicitly.
Claude.ai's account-wide profile instructions (Settings → Profile / Personalization) apply to every chat across the product, and Styles control formatting/tone. The Claude Code analog is `~/.claude/CLAUDE.md`, a user-scoped memory file loaded into every project session on that machine. Use it for stable personal preferences (preferred languages, response style, naming conventions). Edit it directly or use the `#` quick-add syntax inside a session, which Claude will offer to save to user memory.
- Warning: Claude.ai profile instructions follow your account across devices; `~/.claude/CLAUDE.md` is local to one machine and one OS user — copy it manually or check it into a dotfiles repo if you want portability.
- Warning: Claude.ai Styles are a separate orthogonal control (format only); Claude Code has no Styles concept — encode tone preferences in your user CLAUDE.md prose.
- Warning: Org-wide 'Organization instructions' (Team/Enterprise on Claude.ai) map to enterprise-managed CLAUDE.md / managed-settings in Claude Code, not to user CLAUDE.md.
- Warning: User-scope CLAUDE.md applies to all projects on the machine including unrelated ones — keep it short to avoid burning context.
Claude.ai Projects bundle three things: per-project instructions, a knowledge base of uploaded files (PDFs/docs/text), and the chats inside that project. The Claude Code analog is project memory in `./CLAUDE.md` at the repo root (checked into git, shared with collaborators) plus `./CLAUDE.local.md` (gitignored, personal-to-this-checkout). Generate a starter file with `/init`, then iterate. Knowledge files in a Claude.ai Project become: the actual repo files (read by Claude Code on demand via Read/Glob/Grep) and, for static guidance docs, additional markdown files referenced from CLAUDE.md or pulled in with `@path/to/file`.
- Warning: Claude.ai Projects use RAG over uploaded knowledge (paid plans get up to ~10x expansion); Claude Code has no RAG layer — it reads files on demand and you pay context tokens for whatever it loads.
- Warning: Project instructions in claude.ai are a single text box; Claude Code memory layers (enterprise > project > user > local) merge automatically — keep each layer focused.
- Warning: Claude.ai Projects share knowledge across all chats inside the project; Claude Code's CLAUDE.md is loaded into every turn of every session opened in that directory.
- Warning: Recommended CLAUDE.md size is under ~200 lines — knowledge-base PDFs from a Project should not be dumped wholesale into CLAUDE.md; convert to concise markdown or leave as referenced files.
Claude Code is included in every paid consumer Claude plan (Pro $20/mo, Max 5x $100/mo, Max 20x $200/mo, prices May 2026), Team ($20/seat annual or $25 monthly Standard; $100/$125 Premium = 5x usage; 5-seat minimum), and Enterprise. Sign into Claude Code with `/login` using the same Anthropic account you use on claude.ai and your subscription is detected — no separate purchase. Usage limits are shared between the web chat and Claude Code on the same subscription: a heavy day in Claude Code eats into the same weekly quotas as web chat.
- Warning: Free plan does NOT include Claude Code — must be Pro or higher (or API PAYG).
- Warning: If `ANTHROPIC_API_KEY` is set in your environment, Claude Code uses API pay-as-you-go billing instead of your subscription — unset it (or use a different shell) to fall back to plan auth.
- Warning: Max plans have TWO weekly limits (all-model + Sonnet-only) plus a 5-hour session cycle; web chat and Claude Code share these, so check `/usage` and `/status` in-app to see remaining capacity.
- Warning: Team/Enterprise Claude.ai admins can restrict which models the org can use; Claude Code respects those restrictions when you log in with an org account.
- Warning: API keys from console.anthropic.com bill the API account, not the subscription — separate quota, separate invoice.
Claude.ai exposes a small set of in-chat tools the model can call: Web search (enabled via the `+` menu or by asking), the Analysis/Code Execution tool (sandboxed JavaScript/Python in a managed sandbox), and connectors as tools. Claude Code has a much larger tool palette rooted in your real machine: Bash (arbitrary shell commands, gated by permission modes), WebFetch, WebSearch, Read/Write/Edit, plus the IDE MCP server's `executeCode` for Jupyter cells. Migrating: prompts that depended on Claude.ai's sandboxed code execution can run as `python` / `node` invocations through Claude Code's Bash tool against your local interpreters.
- Warning: Claude.ai's Code Execution runs in Anthropic's sandbox (no access to your filesystem); Claude Code's Bash tool runs on your machine and can touch anything in `cwd` — set permissions (`default`/`plan`/`acceptEdits`/`bypassPermissions`) appropriately.
- Warning: Web search semantics differ: Claude.ai uses dynamic filtering via in-search code execution; Claude Code's WebSearch is a tool the agent calls and the results stream into the conversation — no automatic filtering loop.
- Warning: Long-running shells, package installs, and database access are easy in Claude Code Bash but impossible in Claude.ai's sandbox.
- Warning: Conversely, Claude.ai's Analysis tool has a turnkey CSV/Excel parsing experience; Claude Code requires `pandas`/`openpyxl` (or the xlsx skill) to do the same.