Claude daily update: Agent Skills turn workflows into files

Claude Code still has 2.1.185 as the latest verified version, and yesterday's post already covered that waiting-state change. Today's stronger Claude signal is broader: Anthropic's Agent Skills turn recurring workflows into files Claude can discover and load when needed.

An Agent Skill is a small package of instructions, metadata and sometimes scripts, templates or reference material. The point is simple: stop pasting the same operating routine into every chat. Write it down once, let Claude read it at the right moment and keep it as part of the team's workspace.

Source: Agent Skills overview, Anthropic docs and Equipping agents for the real world with Agent Skills, Anthropic Engineering.

Claude daily update: Agent Skills are workflows as files

Anthropic describes Agent Skills as reusable, filesystem-based resources for domain knowledge: workflows, context and best practices. A skill starts with metadata in SKILL.md, such as name and description. When a task matches, Claude can read more of the instruction and, if needed, follow links to extra files or scripts.

The useful idea is progressive loading. Claude does not need to carry the whole manual in context from the start. It can see that a skill exists, open the main file when it matters and read details only when the task needs them.

Source: Agent Skills overview, Anthropic docs.

What this means for practical Claude teams

For Hammer readers, Agent Skills matter most when Claude should do the same kind of work many times: prepare a decision note, analyze a quote, summarize customer cases, build a presentation or check a report against an internal standard.

Anthropic's docs list pre-built API skills for PowerPoint, Excel, Word and PDF. They also say custom skills can carry organizational knowledge and work across several Claude environments, including Claude Code, the Claude API and claude.ai settings.

That makes a skill different from a clever prompt. The prompt is the assignment right now. A skill is the team deciding: "This is how we do this work here."

Source: Get started with Agent Skills in the API, Anthropic docs.

Integrate Skills with enough control

When a skill contains scripts, templates or links to internal sources, it becomes part of the work environment. Treat it that way. Put it under version control if it governs important work, keep the instructions short, and separate read access from write or publishing access.

For real integrations: use environment variables or a secret manager for keys, scope tool access per workflow, redact sensitive output and keep approval gates before Claude writes to customer data, publishes material or changes systems. MCP, Model Context Protocol, is the standard Anthropic uses to connect Claude to external tools and data sources. Skills can describe how the work should be done; MCP can provide controlled access to the tools.

Source: Agent Skills overview, Anthropic docs and MCP tunnels, Anthropic docs.

Try this prompt this week

Human step: collect one routine the team already repeats, two examples of good input, one example of good output and the rules for what Claude may read or change. Put the material where Claude or Claude Code already has approved access.

Read the routine, examples and access rules in this folder.
Propose a minimal Agent Skill for this workflow.
Draft a short SKILL.md with name, description, when to use it and which files it may read.
Mark which steps need human approval before Claude writes, publishes or changes anything.
Keep the proposal short enough for the team to review in ten minutes.

Good output should include:

• one narrow skill for real recurring work
• a clear split between instructions, examples and secrets
• an approval point before writing or publishing
• questions where Claude lacks context instead of guesses

Hammer angle

This is a useful Claude daily update precisely because it is not about another larger model. It is about making the work repeatable. If Claude already helps in chat but the result varies from person to person, do not start with more automation. Start with one small skill: a routine, a few sources, a review point and a simple log of what changed.

That is often where Tool Forge starts. Not with a large platform, but with a workflow the team actually recognizes.