Claude Code skills make repeated AI work easier to control

Claude Code is getting more interesting at the small end, not only in the big agent demos. In 2.1.157, plugins in .claude/skills directories can load automatically without a marketplace, and 2.1.158 makes Auto mode available on Bedrock, Vertex, and Foundry for Opus 4.7 and Opus 4.8 when CLAUDE_CODE_ENABLE_AUTO_MODE=1 is set. That sounds technical. For a normal team, the practical point is simpler: repeated AI instructions can start becoming local, reviewable work routines.

Source: Claude Code CHANGELOG.md

The current signal

Yesterday's signal was about bigger Claude jobs and smaller review packets. Today's useful follow-up is not to launch more agents. It is to give repeated agent work a place where the team can read, change, and version the instructions.

A skill in Claude Code is a reusable instruction that can include workflows, references, scripts, and supporting files. It can be personal, project-scoped, enterprise-scoped, or part of a plugin. Anthropic also describes skills as lighter than putting everything into CLAUDE.md, because they load when needed.

Source: Extend Claude with skills

A plugin is the package around those routines. It can include skills, agents, hooks, MCP servers, LSP servers, background monitors, executables, and default settings. Claude Code's documentation suggests a simple pattern: start with standalone .claude/ configuration for one project, then package it as a plugin when the routine is ready to share.

Source: Create plugins

What small teams should try

Pick a routine that already gets repeated in chat: reviewing a release, summarizing a change, preparing a customer report, checking an import file, or getting a pull request ready. Do not write a huge agent manual. Start with a local skill that says:

• Which files Claude should read first
• Which systems or folders are out of bounds
• What evidence Claude should leave behind
• When a human must approve the next step

This is also a good Tool Forge moment for Hammer Automation: not more AI as the answer, but a real routine with permissions, logs, and a clear stop before anything expensive or sensitive happens.

Do the human setup first

The human does the setup: choose the routine, give Claude read access to the right project material, and check that your Claude Code settings match the team's rules. Claude Code's settings documentation describes managed, user, project, and local scopes, plus permissions, plugins, MCP servers, and environment variables.

Source: Claude Code settings

Then Claude can help draft the first version of the routine.

Try this prompt this week

Inspect this project’s README, .claude folder, settings, scripts, and existing docs.
Find one repeated AI-assisted workflow that should become a local Claude Code skill.
Draft a small SKILL.md for it: trigger, allowed context, steps, evidence to leave, and approval point.
Also list any settings or permissions I should review before using it.
Do not edit files yet. Show the proposed skill first.

The point is not to make Claude freer. It is to make good working habits easier to repeat without every person needing to remember the same long prompt every time.