Gemini Vision (Advisory)

Convert a mockup or screenshot into structured implementation guidance. Do not edit files.

Arguments

• /gemini-vision <image>

  — return implementation guidance inline in terminal output (default)

• /gemini-vision --json-file <image>

  — machine-parseable JSON file flow

Image Syntax

Use file-path references inside the prompt:

gemini -m gemini-3.1-pro-preview --include-directories /tmp -p "Analyze the UI in ./screenshot.png"

Temp File Convention

All temp files use

$PPID

$PPID

$PPID

mktemp -d

$PPID

echo $PPID

Preflight

• Confirm the image path exists:

  ls -la <path/to/image.png>

  .
• Run

  gemini --version

  (or

  gemini -h

  ) to confirm the CLI is available.
• If either command fails, report the error and stop.

Approval-Mode Guardrail

• Preferred mode is

  --approval-mode plan

  for read-only review behavior.
• If Gemini exits with

  Approval mode "plan" is only available when experimental.plan is enabled

  , stop and ask user to enable it in

  ~/.gemini/settings.json

  :

  {
    "experimental": {
      "plan": true
    }
  }
  

• Do not silently fall back to

  --yolo

  . Only use

  yolo

  if the user explicitly asks for that risk tradeoff.

Invocation

JSON-file mode (

--json-file

)

Pre-invocation cleanup (prevents stale reads from prior runs):

rm -f /tmp/gemini-vision-$PPID.json /tmp/gemini-vision-$PPID.err

gemini --approval-mode plan -m gemini-3.1-pro-preview --include-directories /tmp --output-format json -p "PROMPT" > /tmp/gemini-vision-$PPID.json 2>/tmp/gemini-vision-$PPID.err; EC=$?; echo "Exit code: $EC"

Use the Bash tool's

run_in_background

true

timeout

json-file

Before reading the output file, call

TaskOutput

block: true

timeout: 600000

Inline mode (default)

Run in foreground and return stdout directly to the user:

gemini --approval-mode plan -m gemini-3.1-pro-preview --include-directories /tmp --output-format text -p "PROMPT"

Do not redirect stdout in inline mode.

No-Search Mode (If Supported)

If the Gemini CLI supports disabling search (check

gemini -h

--no-search

Prompt Template (strict)

You are converting a UI image into implementation guidance.
This project has a Tauri desktop dashboard for RGBW LED lighting configuration.
Frontend: Svelte 5 (runes syntax) + TypeScript + Tailwind CSS.
Design system: zinc color palette, consistent spacing via Tailwind utility classes.
Key UI areas: device connection, color picker (RGBW/HSV), LED strip configuration, PWM channel control, preset management, serial monitor.
Dashboard location: tools/dashboard/ (Svelte components in src/lib/components/, stores in src/lib/stores/, Tauri backend in src-tauri/src/).
Focus on Svelte component structure and design-system mapping, not pixel-perfect rendering.
Return STRICT JSON only.

OUTPUT JSON SCHEMA (schema-first):
{
  "summary": "one-paragraph visual summary",
  "layout": "structure and hierarchy",
  "components": ["component list — Svelte components mapped to dashboard structure"],
  "tokens": {
    "colors": ["zinc-* Tailwind classes or RGBW hex values where relevant"],
    "typography": ["font families, sizes, weights — use Tailwind classes"],
    "spacing": ["key spacing values using Tailwind spacing scale"]
  },
  "interactions": ["motion/hover/focus behaviors, Tauri command interactions"],
  "accessibility": ["a11y considerations"],
  "implementation_steps": ["ordered, concise steps"],
  "questions": ["missing info to clarify"],
  "confidence": 0.0
}

INTENT SUMMARY:
- <short summary of what to build and why>

BOUNDARIES:
- <what to ignore or exclude>

UNCERTAINTIES:
- <explicit doubts or areas to double-check>

IMAGE (explicit):
- Path: <path/to/image.png>
- Viewport assumptions: <desktop/mobile if known>

CONSTRAINTS:
- Follow zinc color palette and existing Tailwind design patterns.
- Use Svelte 5 runes ($state, $derived, $effect) — no legacy reactive syntax.
- Dashboard communicates with device via Tauri invoke commands.
- Advisory only; do not modify files.

Rules:
- Prefer Tailwind utility classes and zinc-* palette tokens over raw values.
- Map visual elements to existing dashboard components when possible.
- Ask questions when the image is ambiguous.
- Do not ask to write files. Return the full response inline in terminal output.

Inline Prompt Variant (default)

Use this prompt style for inline mode:

You are converting a UI image into implementation guidance.
Return concise, actionable guidance inline in this terminal response.
Include: summary, mapped components, implementation steps, and open questions.
Do not ask to create/write files.

After Invocation

For

--json-file

mode

1. Get the literal PPID value:

   echo $PPID

2. Read

   /tmp/gemini-vision-<PPID>.json

   with the Read tool (using the numeric value)
3. Validate JSON and extract implementation guidance
4. If the file is empty or Gemini exited non-zero, read

   /tmp/gemini-vision-<PPID>.err

   and report the error
5. Post-use cleanup — remove temp files to avoid leaking context:

   rm -f /tmp/gemini-vision-$PPID.json /tmp/gemini-vision-$PPID.err
   

For inline mode

1. Read stdout directly from the command result.
2. If output is missing or command fails, inspect stderr and report the error.

Error Handling

--approval-mode plan

experimental.plan

--yolo

/tmp/gemini-vision-<PPID>.err

TaskOutput

Post-Gemini Triage (Claude+Codex)

1. Validate JSON and extract implementation steps.
2. Map components to existing dashboard patterns:
   • Svelte components:

     tools/dashboard/src/lib/components/

   • Stores:

     tools/dashboard/src/lib/stores/

   • Tauri commands:

     tools/dashboard/src-tauri/src/

3. Confirm tokens against zinc palette and existing Tailwind patterns.
4. Turn open questions into a short clarification list.

Calibration (Important)

• Confirm Gemini can read the provided image path with a small smoke test before relying on output.