SLOPSHOPPER

prompt-enhancer

Rewrites a vague draft in the prompt box using the project's instructions, skills and conversation; Send submits it.

newbandtoastpromptmodel
v0.1.0no licenseupdated 2026-10-09aravindsriraj/claude-code-mods/prompt-enhancer
A shopper browsing a rack in a slop shop
Preview · a replayed session in a sandbox
claude · ~/work/app · prompt-enhancer
› fix the failing auth test and add an audit log call ⏺ Read(src/auth.ts) ⎿ Read 6 lines ⏺ Update(src/auth.ts) ⎿ Added 2 lines, removed 1 line ⏺ Bash(bun test) ⎿ 3 pass, 1 fail ● Done. refresh now rejects expired claims and logs an audit event. ✻ Worked for 42s · done 4:20 PM [ ✨ Enhance ] rewrite the draft with project context ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts

Draws

Band
[ ✨ Enhance ] rewrite the draft with project context
README

claude-code-mods

Four mods for Claude Code:

  • prompt-enhancer turns a rough prompt into a specific one, using what Claude already knows about your project, before you send it.
  • design-picker shows design-direction questions as visual mockups you pick from, instead of a plain list.
  • assumption-tracker has Claude write down each assumption it makes while it works, so you can mark them right or wrong and send corrections back.
  • idea-shelf parks side-ideas mid-task without Claude seeing them, and sends any of them later as a prompt.

Install

Type these at the prompt of a Claude Code session in a terminal:

/plugin install prompt-enhancer --marketplace aravindsriraj/claude-code-mods
/plugin install design-picker --marketplace aravindsriraj/claude-code-mods
/plugin install assumption-tracker --marketplace aravindsriraj/claude-code-mods
/plugin install idea-shelf --marketplace aravindsriraj/claude-code-mods

The first time, Claude Code asks whether to add the marketplace (github:aravindsriraj/claude-code-mods): answer y. Then choose a scope; the user scope is listed first, and Enter picks it. The mod starts working in that session right away, and with the user scope in every session after.

Mods built on Claude Code's hooks API need a recent Claude Code; the API is in early access and may change between releases.

prompt-enhancer

A one-line bar above the prompt box with an ✨ Enhance button.

  1. Type a rough prompt as usual, e.g. fix the export bug.
  2. Press ✨ Enhance: click it (fullscreen terminal), or press ctrl+x tab, then e. The bar shows "Enhancing…", then your draft is replaced with a rewritten prompt.
  3. Read it, edit it if you like, then press Send (s) or just Enter. Undo (u) puts your original draft back.

The rewrite sees what the main model sees: your CLAUDE.md and memory files, the installed skills, and the conversation so far, so "the bug" can become a real file and function. On the first prompt of a new session, before there is a conversation, it reads your instruction files and skill names directly.

Good to know:

  • The rewrite uses your session's model, so it takes a few seconds.
  • The Send button doesn't expand @file mentions or pasted images, and the transcript shows the message as sent by the mod. For those prompts, press Enter instead.
  • The bar is always shown, even when the prompt box is empty.

design-picker

When Claude asks you to choose a design direction with its AskUserQuestion tool (often while using the frontend-design skill), each option is drawn as a mockup above the usual list:

  • Desktop app: an SVG card per option. If the option's preview is an SVG it is shown as is; otherwise the card shows the label, description, color swatches for any hex colors mentioned, and the preview text.
  • Terminal: a Pick button per option with its color swatches, and the first lines of its preview under it.

Press Pick on a mockup, or answer in the list below it as usual; whichever you do first is the answer Claude gets.

Which questions it takes: single-pick questions whose header or text mentions design, direction, aesthetic, visual, theme, palette, colour, typography, layout, mockup, UI or UX, when every question in the call is one. Everything else, including a call that mixes design and other questions, is drawn exactly as before.

Good to know:

  • Claude Code allows 12 rows around this dialog, counting every line, button and image label. That fits one question with up to 4 options on desktop, or two such questions on the terminal. A larger set of questions gets the normal dialog, unchanged.
  • On the terminal, the preview lines share what is left of those rows: 1 line each with 4 options, 4 each with 2. The normal list still shows the full preview of the option you move to.
  • Terminal Pick buttons have no hotkeys, so they don't clash with the dialog's number keys. Click them in a fullscreen terminal, or answer in the list.

assumption-tracker

While Claude works, it records each assumption it acts on: a requirement nobody stated, a project convention, a fact about the code it hasn't checked. It does this by calling the mod's record_assumption tool, which the mod lists in Claude's tools and tells it about in the system prompt, then keeps working without stopping to ask.

  1. Once there is at least one, a line above the prompt shows Assumptions (3), 2 to review [Show].
  2. Press Show: click it, or press ctrl+x tab, then a. That opens a side pane listing every assumption, with Claude's reason when it gave one.
  3. For each, press ✓ Right or ✗ Wrong, or type what is actually true in its Correction field and press Enter (that marks it wrong).
  4. Press Send feedback to send your verdicts to Claude now, as a message from the mod. Or just carry on: verdicts you haven't sent go along with your next prompt, unseen in the transcript. Each verdict is sent once; the pane shows which ones Claude has.

Good to know:

  • How often Claude records depends on the model following the instruction; it is a prompt, not a guarantee.
  • Send feedback starts a new turn once Claude is idle. If Claude is mid-task, sending your next prompt carries the verdicts too.
  • A /clear starts the list over.
  • In the mobile app there is no Correction field: Right and Wrong only.

idea-shelf

Park a side-idea without derailing what Claude is doing, and send it later.

CommandWhat it does
/idea <text>Puts the idea on the shelf with a timestamp. Runs at once, even while Claude is working, and shows a 💡 toast.
/ideasOpens the shelf in a side pane, numbered, with a Send button per idea. Press its number (or click Send) to run it. Esc closes.
/idea-send <n>Sends idea n as a prompt and takes it off the shelf.

A sent idea is submitted as your own prompt (the transcript labels it "Prompt from the idea-shelf plugin"). If Claude is busy it waits its turn.

Claude never sees a shelved idea. The mod replaces the text of your /idea command with [shelved for later; hidden from Claude] before it reaches the conversation, so Claude only learns that something was shelved. /ideas draws the list in a pane, outside the conversation. Your typed text does stay in your local prompt history (up-arrow) and in the session file's record of the command, neither of which Claude reads.

The shelf lives in ~/.claude/idea-shelf.md, one idea per line (- [2026-10-09 14:32] your idea), shared by every project and session. You can edit it by hand.

Development

Each mod is a folder with .claude-plugin/plugin.json, hooks/hooks.json and the hooks module hooks/register.tsx; .claude-plugin/marketplace.json at the root lists them all.

claude plugin validate .
claude plugin validate ./prompt-enhancer
claude plugin test ./prompt-enhancer
claude plugin validate ./design-picker
claude plugin test ./design-picker
claude plugin validate ./assumption-tracker
claude plugin test ./assumption-tracker
claude plugin validate ./idea-shelf
claude plugin test ./idea-shelf

To try a change, start a session with the mod loaded from your working copy: claude --plugin-dir ./prompt-enhancer. Claude Code then writes the API's type declarations into .claude-plugin/types/ and a tsconfig.json beside them, so tsc -p ./prompt-enhancer type-checks it; both are gitignored.

Source 2 files
hooks/register.tsx 111 lines
1import { atom, read, update } from 'claude-code'
2import type { EngineInterface, ModelForkResult, Register } from 'claude-code'
3
4import type { Phase } from '../types'
5
6const phase = atom({ plugin: 'prompt-enhancer', key: 'phase' } as const, 'idle' as Phase)
7const original = atom({ plugin: 'prompt-enhancer', key: 'original' } as const, '')
8
9const ask = (draft: string) => `Rewrite the draft prompt below into a clear, specific prompt for you (Claude Code) to act on in this project.
10Ground it in what you know: this conversation, the project's instructions (CLAUDE.md, rules, memory) and the skills and tools available.
11- Keep the person's intent; do not invent requirements or widen the scope.
12- Resolve vague references ("it", "the bug", "that page") to concrete files, functions or features when the context makes them clear.
13- Name the skill, command or project convention that applies when one clearly does.
14- Say what done looks like (how to verify) when it is not obvious.
15- Be concise: a few sentences or a short list, never an essay.
16Reply with the rewritten prompt only: no preamble, no quotes, no code fence.
17
18<draft>
19${draft}
20</draft>`
21
22// The fork sees exactly what the main model sees: system prompt, CLAUDE.md, skill listing, conversation.
23async function rewrite($: EngineInterface, draft: string): Promise<ModelForkResult> {
24  const forked = await $.model.fork({ prompt: ask(draft) })
25  if (forked.isAnswered || forked.reason !== 'nothing-to-fork') return forked
26
27  // Fresh session, nothing to fork yet: hand over the instruction files and skill names ourselves.
28  const { breakdown } = (await $.session.usage({ breakdown: 'summary' })).context
29  const files = await Promise.all(
30    (breakdown?.memoryFiles ?? []).map(
31      async f => `<file path="${f.path}">\n${await $.fs.read(f.path).catch(() => '')}\n</file>`,
32    ),
33  )
34  const skills = (breakdown?.skills?.skillFrontmatter ?? []).map(s => s.name).join(', ')
35
36  return $.model.complete({
37    model: await $.session.model(),
38    system: `Project instructions:\n${files.join('\n')}\n\nInstalled skills: ${skills || 'none'}`,
39    prompt: ask(draft),
40  })
41}
42
43export const register: Register = on => {
44  // Sent by Enter instead of the band's Send: back to idle.
45  on('prompt.submit', async ($, e, next) => {
46    if ((await read($, phase)) === 'ready') await update($, phase, () => 'idle')
47    return next(e)
48  })
49
50  on('ui.render', { component: 'AbovePrompt' }, async ($, e, next) => {
51    if (e.props.hasSurvey) return next(e)
52
53    const { Box, Button, Text } = $.ui.resolve(e)
54    const now = await read($, phase)
55
56    const enhance = async () => {
57      const { text } = await $.prompt.read()
58      if (!text.trim()) return $.ui.toast('Type a prompt first, then Enhance.')
59
60      await update($, phase, () => 'working')
61      const r = await rewrite($, text)
62      if (!r.isAnswered) {
63        await update($, phase, () => 'idle')
64        return $.ui.toast(`Enhance failed: ${r.reason}`)
65      }
66      await update($, original, () => text)
67      await $.prompt.fill({ text: r.text.trim() })
68      await update($, phase, () => 'ready')
69    }
70
71    const send = async () => {
72      const { text } = await $.prompt.read()
73      if (!text.trim()) return
74      await $.prompt.fill({ text: '' })
75      await update($, phase, () => 'idle')
76      await $.prompt.submit({ text, asUser: true })
77    }
78
79    const undo = async () => {
80      await $.prompt.fill({ text: await read($, original) })
81      await update($, phase, () => 'idle')
82    }
83
84    if (now === 'working') {
85      return (
86        <Box>
87          <Text dimColor>✨ Enhancing with project context…</Text>
88        </Box>
89      )
90    }
91
92    if (now === 'ready') {
93      return (
94        <Box>
95          <Text dimColor>✨ Enhanced, review it below </Text>
96          <Button key="send" label="Send" hotkey="s" variant="primary" onPress={send} />
97          <Text> </Text>
98          <Button key="undo" label="Undo" hotkey="u" onPress={undo} />
99        </Box>
100      )
101    }
102
103    return (
104      <Box>
105        <Button key="enhance" label="✨ Enhance" hotkey="e" onPress={enhance} />
106        <Text dimColor> rewrite the draft with project context</Text>
107      </Box>
108    )
109  })
110}
111
types/index.d.ts 8 lines
1export type Phase = 'idle' | 'working' | 'ready'
2
3declare module 'claude-code' {
4  interface PluginState {
5    'prompt-enhancer': { phase: Phase; original: string }
6  }
7}
8