Why Claude’s 17 New Capabilities Matter: Moving Agents from Chat to Real Workflows

Claude has introduced a set of 17 new capabilities—Projects, Memory, Artifacts, Claude in Chrome, Cowork, Skills, Design, Prompt Caching, Extended Thinking, and others. While a feature list looks like a collection of hidden tricks, the author argues that these functions form a coherent workflow chain that prepares an agent to operate beyond the chat window.

Five‑layer workflow health check

The author proposes evaluating an agent’s readiness with five questions:

Is there a stable context entry (Projects, Memory, Project instructions, CLAUDE.md, Cowork projects)?

Are the outputs auditable artifacts (Artifacts, Files, Design) that can be reviewed and reused?

What real systems can the agent reach, and are the permissions clearly defined (Chrome, Cowork, Connectors, Claude Code)?

Are repeatable processes captured as Skills or Scheduled Tasks instead of being scattered in prompts?

Is reasoning, cost, and permission control managed through Extended Thinking and Prompt Caching?

These five layers map to a concise "AI workflow health‑check":

上下文是否有入口：Projects / Memory / Project instructions / CLAUDE.md
产物是否可审查：Artifacts / Files / Design
触达是否有权限：Chrome / Cowork / Connectors / Claude Code
流程是否可复用：Skills / Scheduled Tasks / workflow rules
成本风险是否可控：Permissions / Incognito / Extended Thinking / Prompt Caching / review

When applied to a real team, the checklist helps identify which layer is missing and what to improve first.

Comparison with Codex

In a previous article about Codex, the focus was on durable threads, goals, and automations—essentially keeping a task running. Claude’s new features complement that by strengthening the surrounding workflow: providing clear context, producing tangible artifacts, and exposing the agent to external systems while demanding explicit permission and cost controls.

Deep dive into each layer

Context layer : Projects act as self‑contained workspaces with chat history, knowledge bases, and uploadable files. Memory and chat search store role, project, and communication preferences, but not everything. CLAUDE.md is a static markdown file that defines commands, code style, project structure, and team conventions; it is read into context each time, not enforced automatically.

Artifact layer : Artifacts turn model output into editable objects—Markdown docs, code snippets, HTML pages, SVGs, charts, React components—displayed beside the chat for immediate iteration. Simon Willison’s use cases (pricing calculators, data tables, interactive prototypes) illustrate how artifacts solve “middle‑ground” problems that are too big for plain text but too small for a full project.

Reach/Permission layer : Claude in Chrome can view and interact with the current webpage; custom connectors can call external services; Cowork accesses local files and desktop apps; Claude Code integrates with terminals, repositories, and CI/CD pipelines. The official docs warn that each expansion of reach must be paired with explicit permission definitions to avoid accidental actions.

Process‑asset layer : Skills are folders containing instructions, scripts, and resources that the agent loads on demand, turning ad‑hoc prompts into reusable, versioned process assets. Scheduled Tasks automate recurring work (daily reports, file organization) but require the computer to stay awake and Claude Desktop to run.

Control layer : Extended Thinking allocates more reasoning budget for complex tasks, while Prompt Caching reduces token cost by reusing stable context prefixes. The API charges cache reads at 0.1× the base input token price, with a default 5‑minute TTL (extendable at higher cost). Properly structuring prompts and context makes caching effective and controls expenses.

Practical rollout guide

Instead of enabling all features at once, the author suggests a step‑by‑step pilot on a real project:

Create a Project and add a concise WORKSPACE.md describing the project, audience, and current goals.

Write a short GOAL.md that states what to achieve, what not to do, acceptance criteria, and pause conditions.

Define a PERMISSIONS.md that classifies files and actions as read‑only, writable, requiring confirmation, or forbidden.

Maintain a SOURCE_INDEX.md that lists official sources, external drafts, historical articles, and pending facts.

Use Artifacts to produce a quick deliverable (comparison table, flowchart, prototype).

When external interaction is needed, enable Chrome, Connectors, Cowork, or Claude Code one at a time, with clear permission scopes.

After three successful runs, extract the stable part into a Skill.

For low‑risk repetitive work, consider a Scheduled Task.

For a newsletter workflow, the same files become WORKSPACE.md (account positioning, reader persona), SOURCE_INDEX.md (raw sources, official docs, curated drafts), GOAL.md (what the article should answer), PERMISSIONS.md (which external references are allowed), and workflows/publish‑check.md (title, front‑page, fact‑check, AI‑tone removal, source citation).

Takeaway

Claude’s usefulness is less about the number of features you collect and more about whether your workflow provides clear entry points for context, auditable outputs, defined reach, reusable processes, and a control surface for permissions, cost, and reasoning. When these layers are in place, the agent can enter, act, and hand off work reliably.

References

Anatoli Kopadze, "Claude Can Do All of This. Most People Have No Idea" (Twitter)

Anthropic Help Center: Projects, Memory, Artifacts, Claude Cowork, Prompt Caching

Simon Willison, "Everything I built with Claude Artifacts this week"

Andrej Karpathy, LLM Wiki gist

Martin Fowler / ThoughtWorks, "Context Engineering for Coding Agents"

Addy Osmani, "agent‑skills" repository