Claude Code 2.0 是一次版本發佈，而非固定能力清單

Claude Code 2.0 was version 2.0 of Anthropic's terminal interface for its coding agent, announced on 29 September 2025. It is not a model, and it is not the product you'd install today.

That distinction matters more than it sounds, because search results blend three things: the historical launch, later 2.0.x package releases, and current behaviour. Teams end up writing rollout policy against a frozen feature list from over a year ago. Use the release for context, then check the current changelog before anything goes into a control document.

Same goes for naming. "Claude Skills 2.0" isn't an Anthropic product name — it's search shorthand. The official terms are Agent Skills (the open standard) and Claude Code skills (the implementation). Useful phrasing for discovery, bad phrasing for an architecture diagram or a procurement record.

What actually shipped in the 2.0 bundle

A refreshed terminal interface with better status visibility and searchable prompt history on Ctrl+R.
A native VS Code extension, in beta at the time. "Beta at launch" is not a current support statement.
Checkpoints, for automatic code-state recovery.
Agent SDK support for subagents and hooks, under the renamed Claude Agent SDK. The durable idea in there is controlled delegation. Everything else has moved.

Checkpoints are local undo, not history

This is the part worth knowing precisely, because it looks like a full undo and isn't. Per the current checkpointing docs:

A checkpoint is created at each prompt that starts a turn, with no configuration.
Claude Code keeps file snapshots for the 100 most recent checkpoints in a session, cleaned up around 30 days after the session's last snapshot (configurable via cleanupPeriodDays).
/rewind, or double-Esc on an empty prompt, opens the menu. You can restore code, conversation, or both.
What it does not cover:
Files modified by bash commands. rm, mv, cp — the docs are explicit. If Claude runs a script that rewrites 300 files, rewind won't bring them back.
External changes. Edits you made outside the session, or from a concurrent session.
Subagent edits, symlinked and hard-linked files. A restore will tell you it skipped them.
Anthropic's own docs put it plainly: treat checkpoints as local undo and Git as permanent history. The practical consequence is a routine rather than a preference — commit before you delegate anything wide-scale, because the more shell autonomy you grant, the more of the work happens outside the checkpoint's reach.

Keep facts in CLAUDE.md, procedures in skills

A skill is a directory built around SKILL.md. Claude loads it when the description matches the task, or you invoke it as /skill-name. The useful split:

CLAUDE.md — persistent project facts that should be available broadly.
Skills — procedures. A code-review checklist, a release-note workflow, a migration playbook. The full body only loads when used, which keeps irrelevant context out and gives you a versioned artifact to review.
Hooks — deterministic actions at defined events, like running tests.
Subagents — isolation or parallelism for specialized work. Checkpoints manage session state. Skills define a procedure. Those are different layers and conflating them is how teams end up with neither.

A skill is code plus policy

Skills don't make a process deterministic. The model still interprets the instructions, the tools still change state, and a third-party skill can carry scripts and references nobody on your team has read. A catalog listing is not a security review.

Six things before one gets installed:

Inspect SKILL.md and every bundled script.
Pin the repository revision or package version.
Limit permissions to the files, commands, networks, and tools the task actually needs.
Test in an isolated repository with representative failure cases.
Record the owner, approved use, review date, and rollback path.
Require a human merge decision for production code. Anything that can run commands, call MCP tools, or fetch network content expands the same attack surface as any other automation dependency — because that's what it is.

Store project skills with the repository when the procedure belongs to one codebase. Use centrally governed distribution when the same approved procedure spans several. Either way it needs a named owner and a review cadence.

Where the gates are worth it, and where they aren't

Sandboxing, managed settings, deny rules, and monitoring reduce risk rather than remove it. Prompt injection, broad tool access, and weak review can still produce a supply-chain or production problem.

But a governed skill takes longer to approve than a private prompt, and that overhead is only justified where it buys something. Strong gates belong on workflows touching production, customer data, authentication, payments, or regulated records. Apply the same gates to every low-risk local task and you get review theatre plus usage quietly migrating to personal accounts — which is worse than where you started, because you've lost the visibility too.

The controls that carry the load are the boring ones: permissions enforced outside the prompt, provenance on the commit, a human signing every merge, Git as the durable source of truth, and a quarterly pass over the changelog to check your controls still match the product.

Cursor or Claude Code

Briefly, because this isn't the comparison article: Cursor is editor-first and keeps the agent inside an interactive coding session. Claude Code started terminal-led and now spans terminal, IDE, desktop, and browser. Cursor usually fits teams who want minimal workflow change; Claude Code fits terminal-heavy and platform work and larger delegated tasks.

Neither product's checkpoints replace Git. Both need ownership and version control on their reusable instructions. Most organizations will end up splitting by task rather than picking a winner.

Full write-up — the release table, the governance control matrix, and where to source skills: teamvoy.com/blog/claude-code-2-0

Written by BYLINE_TO_CONFIRM, Teamvoy. More engineering writing at teamvoy.com/blog.