Fine Print

A second reader that flags what you would miss in a wall of text.

Catch what you would missv0.1.0

What it can touch

Files `none`

Does not read or write files.

Shell commands `none`

Does not run commands.

Network `none`

Does not make network requests.

Model calls `allowed`

Can call a model.

Sound `none`

Does not play sound.

Hooks it can block `none`

Does not use hooks.

Reported by Claude Code's own validator.

The prompt

fine-print / prompt

Build me a Claude Code mod called "fine-print": a second reader that flags what I might miss in long replies.

What it does

- After a turn with a long reply (over about 150 words), ask a small, cheap model to read the reply and a summary of that turn's tool calls, and decide whether there is anything I must know: hidden costs, security or privacy risk, data loss, irreversible actions, impact on customers, or assumptions Claude made without asking.

- If there is, show a note above my prompt: a bold one-line headline and up to 4 bullets, with buttons Got it, Explain more (sends a follow-up question into the chat) and Dismiss.

- If there is nothing important, stay silent. Silence is the default.

What is new here

- A watchFor setting to choose the topics: cost, security, data-loss, irreversible, customers, assumptions. Default: all.

- A spend cap: maxChecksPerSession (default 30). When it is reached, stop and say so once in the status line.

- /fineprint lists every note flagged this session, newest first, so nothing is lost after a dismiss.

- Use the cheapest model available and send only the reply plus tool summaries, never whole files.

Build rules

- Mods are new (Claude Code 2.1.287+). Use your plugin-authoring skill and read the types it points to before writing anything.

- Only do what is described here: no network calls and no reading files the mod does not need, unless this prompt says so.

Details

Fine Print

A second reader that flags what you would miss in a wall of text.

After a long reply, a small, cheap model reads it with a summary of the tool calls behind it. If there is something you must know (a hidden cost, a security or privacy risk, data loss, an action that cannot be undone, an effect on customers, or an assumption made without asking), a short note appears above your prompt. Otherwise it stays silent.

Use it

There is nothing to run. After any reply over 150 words, Fine Print checks it in the background, after the turn has ended.

When it finds something, a band appears above the prompt:

a bold one-line headline and up to 4 bullets
Got it (g): you read it; the band closes
Explain more (e): sends a follow-up question into the chat, so Claude explains the note in plain words. The question is sent as the Fine Print plugin's message, not as yours: the transcript and the model both see that the plugin wrote it, at your press.
Dismiss (d): closes the band without marking it read

The letter keys work once the band has focus (a click, or ctrl+x tab). A newer note replaces the one on screen; the band then says how many older notes are still open.

/fineprint lists every note flagged this session, newest first, with what you did with it (open, got it, dismissed, asked to explain), how long ago, its topics, and how many checks were used.

Settings

Change them in /config (or on the install screen).

Setting	Default	What it does
`watchFor`	`cost, security, data-loss, irreversible, customers, assumptions`	The topics to flag. A list; `all` means every topic, `none` turns checks off. Names that match no topic are ignored, and if none are left every topic is watched, so a typo never switches the reader off.
`maxChecksPerSession`	`30`	The spend cap: model checks allowed per session (0 to 1000). When it is reached, checks stop and the status line says so once. `0` turns checks off.

What it can touch

Model calls: one completion on the haiku model per reply over 150 words, at low effort, with at most 400 output tokens and a 30 second time limit, and never more than maxChecksPerSession per session. It goes through your session's own client and provider. A failed call counts toward the cap.
What is sent: the reply (whole up to 12,000 characters, beyond that its first 8,000 and last 4,000) and up to 40 lines naming this turn's tool calls and their targets: the first line of a command (up to 160 characters), a file path, a URL, a search pattern, and similar. Never file contents, edit text or tool output.
Reads: the session transcript, through Claude Code, to find this turn's tool calls. No files, no processes, no network of its own.
Writes: nothing on disk. Notes and the check count live in the session and end with it.
Explain more starts one turn of your main model, at its normal cost, and only when you press it.
/fineprint prints its list into the transcript, where the model can read it too.

Limits

A small model's judgment: it can miss things and can flag noise. It is a second reader, not a guarantee.
Only main-conversation replies that finished normally are checked: not subagents, not interrupted or failed turns.
"This turn's tool calls" are the calls since your last prompt. A row Claude Code adds that looks like a prompt can shorten that list.
The band draws in the terminal and the desktop app. In other surfaces (VS Code, mobile) a note arrives as a toast instead, and /fineprint has the details.
The note appears a few seconds after the turn ends, while the small model reads.
If the model cannot be reached (for example, an organization policy blocks haiku), Fine Print shows one toast per session and otherwise stays quiet.
The status line about the cap stays up for the rest of the session, unless you raise the cap.

Built from a prompt

This mod was built from prompts/fine-print.md.

Pairs well with

last-lookLast Lookant-farmAnt FarmhatchlingHatchling

What it can touch

Files none

Shell commands none

Network none

Model calls allowed

Sound none

Hooks it can block none

The prompt

Details

Fine Print

Use it

Settings

What it can touch

Limits

Built from a prompt

Pairs well with

Fine Print

Use it

Settings

What it can touch

Limits

Built from a prompt

Files `none`

Shell commands `none`

Network `none`

Model calls `allowed`

Sound `none`

Hooks it can block `none`