Claude Code mods in practice: a cache clock, a record-mode blur, goal progress and a file-collision guard

What a mod is (checked against Anthropic)

Anthropic's mods overview defines a mod as a plugin that changes how Claude Code looks and behaves, made of JavaScript or TypeScript event handlers that run inside Claude Code. Mods work in the Claude Code CLI and the Code tab of the Claude Desktop app, and you can get one by describing what you want in a session and letting Claude write it. Nate's description matches: mods ship inside plugins, you manage them with /plugin, and you can ask Claude to build one in plain language. They arrived in Claude Code 2.1.287 (see our changelog entry).

The four mods he demonstrated

1. Cache keeper. A panel shows context size (109,000 tokens in his example), the 5-hour and weekly limits, and what the session would have cost at API prices. It also tracks the prompt cache: his display counts down from 60 minutes and, because an expired cache means repaying for the whole conversation, it sends a notification shortly before the cache goes cold, offering to re-prompt, compact or hand off. A button runs his session-handoff skill and then /clear, pasting the handoff message into the fresh session, which turns copy, clear, paste into two clicks.
2. Record mode. While on, emails, financial figures and other private text are replaced with placeholders on screen. A record command turns it on and an off argument turns it off. Claude still receives the real text; in his demo a price change from $29 to $39 was applied correctly while the display stayed masked.
3. Goal progress. Shows a small progress view when you start a /goal prompt: elapsed time, the task list Claude builds (four tasks in his example) and a percentage complete, with one place to see goals running in several sessions.
4. Collision guard. Before any edit, write or notebook edit, it checks whether the target file was touched in the last 30 minutes. If so it asks in the interface: proceed, move to a worktree, cancel, or type something else. He points out this differs from a pre-tool-use hook, which blocks or acts automatically without asking.

How he built them

He asked Claude to read his session logs and propose the five mods most likely to save him money and time, took ideas from that, and installs a finished mod by pasting its text into Claude with "set up this mod". He notes you can also feed his video transcript to your agent and describe the behaviour you want. The mods he built are distributed through his free community, which this page does not reproduce.

Gotchas & Caveats

• The 60-minute cache window is the figure his panel shows. Anthropic's API documentation lists a 5-minute default prompt-cache lifetime with a 1-hour option at higher cost, so check what applies to your plan before relying on any countdown.
• A mod runs code inside Claude Code. Read a mod before enabling it, especially one received from someone else.
• Record mode masks the screen, not the request: Claude is still sent the real values.
• The collision guard is a prompt in the interface, not a block, so an unattended session will not be protected by it.
• Several mods in the video are demos for his own workflow (he calls one, a virtual pet, 'pretty useless'). The video also contains an unrelated plug for a free SOP.

Key Takeaways

• Mods are the new way to add behaviour that settings hooks and skills cannot reach, because they run inside Claude Code and can draw interface.
• The most transferable idea is the cache clock: warn before a long idle session turns expensive, then offer a handoff.
• Ask Claude to mine your own session logs for repeated actions before deciding which mods to build.
• A mod that asks in the UI (collision guard) complements, rather than replaces, a hook that blocks.