NothingFancyAI · wesnick · May 2, 2026 · May 2, 2026 · May 2, 2026 · May 2, 2026
@@ -53,6 +53,17 @@
         "url": "https://nothingfancy.ai"
       },
       "source": "./plugins/gemini-tts-prompting"
+    },
+    {
+      "name": "launching-google-ai",
+      "version": "0.1.0",
+      "description": "Opens Google Gemini or Google Stitch in the browser with a pre-filled prompt and optional Gemini tool (image, video, music, deep research, canvas, guided learning). Use when asked to 'ask Gemini', 'open Gemini with', 'stitch this', 'use deep research on', 'generate a video in Gemini', or otherwise launch a prefilled Gemini or Stitch session via URL parameters from the shell. Does not auto-submit — only pre-fills the prompt box.",
+      "author": {
+        "name": "Wesley O. Nichols",
+        "email": "wes@nothingfancy.ai",
+        "url": "https://nothingfancy.ai"
+      },
+      "source": "./plugins/launching-google-ai"
     }
   ]
 }
@@ -0,0 +1 @@
+../../plugins/launching-google-ai/skills/launching-google-ai
@@ -27,6 +27,6 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      - uses: astral-sh/ruff-action@v1
+      - uses: astral-sh/ruff-action@v3
         with:
-          version: latest
+          args: check
@@ -20,3 +20,5 @@ evals/transcripts/*
 evals/reports/*
 !evals/reports/.gitkeep
 evals/.promptfoo-cache/
+
+scratchpad/
@@ -9,6 +9,7 @@
 /plugins/justfile/                         @wesnick
 /plugins/nano-banana-prompting/            @wesnick
 /plugins/gemini-tts-prompting/             @wesnick
+/plugins/launching-google-ai/              @wesnick
 
 # Shared infrastructure
 /.github/                                  @wesnick

@@ -34,6 +34,7 @@ This is the preferred mode when you want a narrow, focused session.
 |--------|-------------|
 | [nano-banana-prompting](plugins/nano-banana-prompting/) | Crafts and reviews prompts for Google's Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) image generation and editing models — frameworks, camera/lighting direction, and text rendering. |
 | [gemini-tts-prompting](plugins/gemini-tts-prompting/) | Crafts and reviews prompts for Google's Gemini 3.1 Flash TTS using audio tags, voice style instructions, and pacing controls — narration, audiobooks, IVR, accessibility, multilingual content. |
+| [launching-google-ai](plugins/launching-google-ai/) | Opens Google Gemini or Google Stitch in the browser with a pre-filled prompt and optional Gemini tool selection (image, video, music, deep research, canvas, guided learning) via URL parameters. |
 
 ## Philosophy
 

@@ -0,0 +1,135 @@
+# Regression suite for launching-google-ai skill
+
+description: "launching-google-ai skill regression cases"
+
+tests:
+  # --- Positive: Gemini deep-research launch ---
+  - description: "Positive — 'ask gemini to research X' should build a Gemini deep-research URL"
+    vars:
+      prompt: |
+        ask gemini to research the history of LISP
+    assert:
+      - type: contains
+        value: "https://gemini.google.com/?prompt="
+      - type: contains
+        value: "tool=research"
+      - type: contains-any
+        value:
+          - "the%20history%20of%20LISP"
+          - "the+history+of+LISP"
+      - type: contains-any
+        value:
+          - "${BROWSER:-xdg-open}"
+          - "$BROWSER"
+      - type: llm-rubric
+        value: |
+          The response must: (1) build a URL on gemini.google.com (NOT
+          stitch.withgoogle.com, NOT any other Google surface), (2) use
+          ?prompt= as the parameter (NOT ?q= or ?query=), (3) percent-encode
+          the prompt text, (4) include &tool=research because the user said
+          "research", using the SHORT alias "research" rather than the human
+          label "Deep research", and (5) open the URL via ${BROWSER:-xdg-open}
+          (or equivalent fallback) so a user-configured browser is honored.
+          The response must NOT attempt to auto-submit the prompt — pre-fill
+          only.
+
+  # --- Positive: Stitch design launch (no tool param) ---
+  - description: "Positive — 'stitch a pricing page' should build a Stitch URL with no tool parameter"
+    vars:
+      prompt: |
+        stitch a clean SaaS pricing page with three tiers
+    assert:
+      - type: contains
+        value: "https://stitch.withgoogle.com/?prompt="
+      - type: not-contains
+        value: "tool="
+      - type: llm-rubric
+        value: |
+          The response must: (1) target stitch.withgoogle.com, NOT
+          gemini.google.com, because the user explicitly said "stitch",
+          (2) NOT include any &tool=... parameter — Stitch does not accept a
+          tool parameter and adding one is wrong, (3) percent-encode the
+          design brief into the prompt= parameter verbatim (the user's words,
+          not paraphrased), (4) open via ${BROWSER:-xdg-open} or equivalent.
+
+  # --- Negative: prompt-craft request should NOT fire this skill ---
+  - description: "Negative — 'help me write a prompt for Gemini image gen' is a prompt-craft task, not a launch task"
+    vars:
+      prompt: |
+        Help me write a prompt for Gemini image generation of a vintage
+        bookstore at golden hour with reference photos I'll attach.
+    assert:
+      - type: not-contains
+        value: "gemini.google.com/?prompt="
+      - type: not-contains
+        value: "xdg-open"
+      - type: not-contains
+        value: "${BROWSER"
+      - type: llm-rubric
+        value: |
+          The response must address prompt-craft for Gemini image generation
+          on its own terms (subject + action + location + composition + style,
+          camera/lighting direction, treatment of reference images, etc.) —
+          this is the territory of the nano-banana-prompting skill, NOT
+          launching-google-ai. The response must NOT construct a
+          gemini.google.com URL, must NOT call xdg-open or $BROWSER, and must
+          NOT instruct the user to launch the Gemini web UI. The launching
+          skill should not fire because the user is asking how to write the
+          prompt, not how to open Gemini.
+
+  # --- Edge: multi-line prompt with special chars must encode newlines ---
+  - description: "Edge — multi-line prompt should encode newlines as %0A and preserve special chars"
+    vars:
+      prompt: |
+        Open Gemini with this multi-line prompt and use deep research:
+
+        Compare Rust and Zig for systems work.
+        Cover: memory model, async story, & C interop.
+        Cite sources.
+    assert:
+      - type: contains
+        value: "https://gemini.google.com/?prompt="
+      - type: contains
+        value: "tool=research"
+      - type: contains
+        value: "%0A"
+      - type: contains-any
+        value:
+          - "%26"
+          - "&amp;"
+      - type: llm-rubric
+        value: |
+          The response must: (1) percent-encode every newline in the
+          multi-line prompt as %0A so the prompt arrives in the Gemini box
+          with line breaks intact, (2) percent-encode the literal "&"
+          character inside the prompt body as %26 so it is NOT interpreted as
+          a URL parameter separator (this is the historically-broken case —
+          an unencoded "&" silently truncates the prompt at the ampersand),
+          (3) still produce a single well-formed URL with ?prompt=... and
+          &tool=research, (4) open via ${BROWSER:-xdg-open}.
+
+  # --- Edge: ambiguous tool cue should default to omitting tool= ---
+  - description: "Edge — ambiguous request without a clear tool cue should omit the tool parameter"
+    vars:
+      prompt: |
+        open gemini and ask it what the weather will be like in Madrid this
+        weekend
+    assert:
+      - type: contains
+        value: "https://gemini.google.com/?prompt="
+      - type: not-contains
+        value: "tool=research"
+      - type: not-contains
+        value: "tool=image"
+      - type: not-contains
+        value: "tool=video"
+      - type: llm-rubric
+        value: |
+          The response must build a Gemini URL with ?prompt=... but must NOT
+          include any &tool=... parameter, because the user's request ("ask
+          it what the weather will be like") does not clearly match any of
+          the six known tool aliases (image, video, music, research, canvas,
+          learn). The skill's documented behavior is to omit tool= when the
+          intent is ambiguous and let Gemini pick a default. The response
+          must NOT fabricate an unsupported tool name like "weather",
+          "search", or "ask".
@@ -0,0 +1,10 @@
+{
+  "name": "launching-google-ai",
+  "version": "0.1.0",
+  "description": "Opens Google Gemini or Google Stitch in the browser with a pre-filled prompt and optional Gemini tool (image, video, music, deep research, canvas, guided learning). Use when asked to 'ask Gemini', 'open Gemini with', 'stitch this', 'use deep research on', 'generate a video in Gemini', or otherwise launch a prefilled Gemini or Stitch session via URL parameters from the shell. Does not auto-submit — only pre-fills the prompt box.",
+  "author": {
+    "name": "Wesley O. Nichols",
+    "email": "wes@nothingfancy.ai",
+    "url": "https://nothingfancy.ai"
+  }
+}
@@ -0,0 +1,52 @@
+# launching-google-ai
+
+Opens Google Gemini or Google Stitch in the browser with a pre-filled prompt and optional Gemini tool selection (image, video, music, deep research, canvas, guided learning).
+
+## What it does
+
+- Builds the right URL for Gemini (`https://gemini.google.com/?prompt=...&tool=...`) or Stitch (`https://stitch.withgoogle.com/?prompt=...`).
+- Picks the correct Gemini tool alias from natural-language cues — "research this" → `tool=research`, "make a video of" → `tool=video`, "generate an image of" → `tool=image`.
+- URL-encodes the prompt safely, including newlines (`%0A`) and special characters.
+- Honors `$BROWSER` (so users can route to a specific Chrome profile or browser binary), falling back to `xdg-open`.
+- Stops at pre-fill — never auto-submits, so the user keeps a final-edit step.
+
+## Usage
+
+### Claude Code
+
+```sh
+claude --plugin-dir ./plugins/launching-google-ai
+```
+
+Or via the marketplace, once installed.
+
+### pi
+
+```sh
+pi -e ./plugins/launching-google-ai    # ephemeral
+pi install ./plugins/launching-google-ai
+```
+
+### Codex
+
+See the root [`.codex/INSTALL.md`](../../.codex/INSTALL.md).
+
+## Harness support
+
+| Feature | Claude Code | Codex | pi |
+|---|---|---|---|
+| Skill (SKILL.md) | ✅ | ✅ | ✅ |
+
+Pure skill content — runs identically in all three harnesses. The only host requirement is a shell with `$BROWSER` or `xdg-open` available (Linux/macOS); on macOS, `open` works as a drop-in replacement for `xdg-open` and the skill will use it via `$BROWSER` if set.
+
+## Evals
+
+Regression cases live in [`../../evals/cases/launching-google-ai/cases.yaml`](../../evals/cases/launching-google-ai/cases.yaml).
+
+```sh
+uv run evals/framework/run.py --skill launching-google-ai
+```
+
+## License
+
+[CC BY-SA 4.0](../../LICENSE) — made by [Nothing Fancy](https://nothingfancy.ai).
@@ -0,0 +1,26 @@
+{
+  "name": "@nothingfancy/launching-google-ai",
+  "version": "0.1.0",
+  "description": "Opens Google Gemini or Google Stitch in the browser with a pre-filled prompt and optional Gemini tool selection.",
+  "keywords": [
+    "pi-package",
+    "agent-skill",
+    "gemini",
+    "stitch",
+    "google-ai",
+    "browser-launcher",
+    "url-prompt"
+  ],
+  "author": {
+    "name": "Wesley O. Nichols",
+    "email": "wes@nothingfancy.ai",
+    "url": "https://nothingfancy.ai"
+  },
+  "license": "CC-BY-SA-4.0",
+  "pi": {
+    "skills": ["./skills"]
+  },
+  "peerDependencies": {
+    "@mariozechner/pi-coding-agent": "*"
+  }
+}
@@ -0,0 +1,109 @@
+---
+name: launching-google-ai
+description: Opens Google Gemini or Google Stitch in the browser with a pre-filled prompt and optional Gemini tool (image, video, music, deep research, canvas, guided learning). Use when asked to "ask Gemini", "open Gemini with", "stitch this", "use deep research on", "generate a video in Gemini", or otherwise launch a prefilled Gemini or Stitch session via URL parameters from the shell. Does not auto-submit — only pre-fills the prompt box.
+---
+
+# Launching Google AI
+
+Opens a Google Gemini or Google Stitch session in the user's browser with a prompt already typed in the box and (for Gemini) a tool already selected. The skill exists because the user wants to hand off to a hosted UI for the next step — usually a multi-modal session, a deep-research run, or a Stitch design loop — without retyping context.
+
+The guiding principle: **build the URL, open it, stop.** Don't auto-submit. The user owns the final edit and the send button, every time.
+
+## When to Use
+
+- The user says "ask Gemini …", "open Gemini with …", "in Gemini, …", "fire up Gemini for …".
+- The user says "stitch …", "stitch this …", "open Stitch with …", or names a UI mockup task they want Stitch to handle.
+- The user wants a specific Gemini tool: "deep research X", "use canvas to …", "generate a video of …", "generate an image of …", "make some music about …", "guided learning for …".
+- The user wants to take prompt text already in the conversation and continue it in a hosted Google AI UI.
+
+## When NOT to Use
+
+- The user wants to **call the Gemini API or Vertex AI** programmatically — that's an SDK task, not a URL launcher.
+- The user wants to **write or refine** a prompt for Gemini image generation — use `nano-banana-prompting`.
+- The user wants to **write or refine** a prompt for Gemini TTS — use `gemini-tts-prompting`.
+- The user wants the model's actual answer in this conversation — answer it here; don't bounce them to the web UI.
+- The user wants Google AI Studio, NotebookLM, or another Google surface — those are not covered (Gemini and Stitch only).
+- The user wants the browser to **submit** automatically — that's outside this skill's contract and the underlying URL parameters do not support it.
+
+## URL Patterns
+
+### Gemini (default)
+
+```
+https://gemini.google.com/?prompt=<encoded_text>
+https://gemini.google.com/?prompt=<encoded_text>&tool=<tool_alias>
+```
+
+### Stitch
+
+```
+https://stitch.withgoogle.com/?prompt=<encoded_text>
+```
+
+Stitch does not accept a `tool` parameter. Adding one is a silent no-op at best; treat it as a hard error and omit it.
+
+## Gemini Tool Aliases
+
+Pick the alias from natural-language cues. The aliases on the right are what go in `&tool=<alias>` — use the short form, not the human label.
+
+| User cue | Tool | Alias |
+|---|---|---|
+| "generate an image", "draw", "make a picture" | Create image | `image` |
+| "make a video", "generate a video", "animate" | Create video | `video` |
+| "write a song", "compose music", "make music" | Create music | `music` |
+| "deep research", "research this thoroughly", "comprehensive report on" | Deep research | `research` |
+| "open canvas", "iterate in canvas", "use canvas for …" | Canvas | `canvas` |
+| "teach me", "guided learning", "walk me through learning …" | Guided learning | `learn` |
+
+If the user's intent doesn't clearly match one of these, omit the `tool` parameter — Gemini will pick a sensible default.
+
+## Building the URL
+
+1. **Pick the surface.** Default to Gemini. Switch to Stitch only when the user explicitly says Stitch or describes a UI/mockup design task that fits Stitch's scope.
+2. **Pick the tool** (Gemini only) using the alias table above. If unsure, omit it.
+3. **URL-encode the prompt** using standard percent-encoding. Newlines become `%0A`. Spaces can be `%20` or `+` (prefer `%20` for readability). Pass raw text — the receiving page handles HTML escaping.
+4. **Assemble**: `?prompt=<encoded>` first, then `&tool=<alias>` if present.
+
+## Opening the URL
+
+```sh
+${BROWSER:-xdg-open} "<url>"
+```
+
+Honor `$BROWSER` so the user can route to a specific browser or profile. Common pattern:
+
+```sh
+export BROWSER="google-chrome --profile-directory='Profile 1'"
+```
+
+On macOS, users typically set `$BROWSER` to `open` or to a specific app bundle path. Don't hardcode `xdg-open` without the `${BROWSER:-…}` fallback — it breaks Mac users.
+
+## Worked Example
+
+User: *"ask gemini to research the history of LISP"*
+
+1. Surface: Gemini (default).
+2. Tool: "research" → `research`.
+3. Encode prompt `the history of LISP` → `the%20history%20of%20LISP`.
+4. URL: `https://gemini.google.com/?prompt=the%20history%20of%20LISP&tool=research`.
+5. Open: `${BROWSER:-xdg-open} "https://gemini.google.com/?prompt=the%20history%20of%20LISP&tool=research"`.
+
+User: *"stitch a clean SaaS pricing page with three tiers"*
+
+1. Surface: Stitch (explicit).
+2. Tool: omit (Stitch does not support `tool`).
+3. Encode prompt `a clean SaaS pricing page with three tiers` → `a%20clean%20SaaS%20pricing%20page%20with%20three%20tiers`.
+4. URL: `https://stitch.withgoogle.com/?prompt=a%20clean%20SaaS%20pricing%20page%20with%20three%20tiers`.
+5. Open with `${BROWSER:-xdg-open}`.
+
+## Anti-patterns
+
+| Anti-pattern | Signal | Fix |
+|---|---|---|
+| Auto-submitting the prompt | The skill tries to script a click or send-key after opening the URL | Pre-fill is the contract. Stop at the URL — the user submits. |
+| Using `?q=` instead of `?prompt=` | URL has `?q=…` | The Gemini/Stitch URL contract is `prompt=`. `q=` does not pre-fill. |
+| Adding `tool=` to a Stitch URL | URL is `stitch.withgoogle.com/?prompt=…&tool=…` | Drop the `tool=` — Stitch ignores it (and "ignore" is a generous read). |
+| Hardcoding the browser | `xdg-open` or `chromium` literal in the command | Use `${BROWSER:-xdg-open}` — Mac and per-profile users depend on it. |
+| Translating tool names | `&tool=Create%20image` | Use the short alias (`image`, `video`, `research`, etc.), not the human label. |
+| Re-typing prompt content into the URL | Prompt in URL diverges from what the user said | Encode the user's prompt verbatim; only normalize whitespace. |
+| Fabricating an unsupported tool | `&tool=summarize`, `&tool=translate`, etc. | Only the six aliases in the table are real. If the user wants something else, omit `tool=` entirely. |
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		../../plugins/launching-google-ai/skills/launching-google-ai