Start with one useful workflow.
You do not need an API key, Google account, or programming experience to run this example. The browser uses a fixed authored response. It does not contact Gemini.
Quick start: your first successful run
- Open the studio and select Start your first AI application.
- Read the five input rows. Each has a numeric ID and message.
- Read the allowed category list. Each category occupies one line.
- Read the authored response. This is a prepared example, not a generated answer.
- Select Run pipeline. Expect five updated preview rows and no errors.
- Select Load invalid category, then Run pipeline. Expect an invalid-category error and no updated rows.
- In Response, change Urgent to Account support. Run again. Expect success.
- Change a message. Run again. Notice that the response does not change. Explain why human review is needed.
- Select Run 10 contract tests. Expect 10 passes. The final case shows that structural validation cannot detect a wrong meaning.
- Select Save locally. Then export a backup to retain a portable copy.
- Open Projects, inspect Code.gs, and download the complete project.
What each area teaches
- Start Here
- Explains Gemini, Apps Script, the fixture boundary and the six-step first activity.
- Learn
- One complete lesson about parsing, contracts, row identity and human review. Record your prediction and reflection. Use hints only after an attempt.
- Build Studio
- Actual JSON parsing, validation, ID matching and local spreadsheet preview. No remote spreadsheet is modified.
- Automation
- Estimate labor savings, review burden and break-even using your own assumptions.
- Projects
- Inspect and download the complete Apps Script source and setup instructions.
- Showcase
- Five-minute presenter rehearsal and downloadable script.
- Help
- This consolidated guide. Navigation stays clear of individual help icons.
Every control in the workbench
- Learning mode
- Beginner shows extra instructions. Guided Builder uses the same tools. Advanced Developer hides beginner guidance. Edits stay intact.
- Save locally
- Explicitly saves your current draft to this browser's IndexedDB. Clearing browser data removes it. Private browsing may not retain it.
- Export backup / Import backup
- Download or restore a version 1 JSON workspace. Import replaces the current in-memory draft, not the saved draft until you select Save locally. Export before importing if you need both versions.
- Inputs / Categories / Response
- Edit source rows, the allowlist and the proposed answer. They are separate inputs. Never paste secrets.
- Inspect constructed prompt
- Shows the actual prompt made from the current rows and categories. It makes no API request.
- Run pipeline
- Parses JSON, validates all IDs and categories, and joins valid candidates to rows. Errors block the entire local result.
- Load valid fixture
- Replaces Response with the original five authored answers. It does not reset edited input.
- Load invalid category / Load malformed JSON
- Replace Response with a prepared failure. Run the pipeline to observe it.
- Run 10 contract tests
- Tests the built-in validator against known positive and negative cases. These are local contract tests, not a Google integration test.
- Inspect actual validation source
- Read the exact function the workbench uses. Edit the exported Code.gs to extend it. This release does not execute arbitrary edited JavaScript.
- Prompt purpose / Build prompt / Copy prompt
- Assemble an editable AI help request with your real code and current data, then copy it manually. No AI service is called.
- Calculate impact
- Computes weekly, monthly and annual hours, estimated financial value and break-even. All time entries are minutes except maintenance and implementation, which are hours.
- Inspect exported file
- Choose a file to view. Export includes complete source, manifest, data and setup instructions.
- Download complete Apps Script project
- Creates a ZIP from your current input and categories. It does not deploy anything.
- Project notes / Prediction / Reflection
- Write your reasoning and extension plan. Save locally or export a backup to keep them.
Get finished code into Google
- Download the project ZIP and extract it.
- Create a new disposable spreadsheet. Open Extensions → Apps Script.
- Replace the default script with Code.gs. Enable the manifest in Project Settings and copy appsscript.json.
- Add SPREADSHEET_ID, GEMINI_API_KEY and GEMINI_MODEL under Script Properties. Keep keys out of browser files and backups. Project editors can access these properties.
- Select a currently supported model that accepts generateContent. Verify account availability and costs in Google AI Studio.
- Run install. Authorize scopes only after reviewing the source. The script refuses to overwrite existing Messages or Review sheets.
- For an offline rehearsal, enter the exact input snapshot JSON into Review!A2, candidate classification JSON into B2, and TRUE into C2 after reviewing every answer.
- For a real request, explicitly run generateDraft. It sends the Messages data to Google and may incur charges. Check the response and its meaning before entering TRUE in Review!C2.
- Run writeApproved. Categories go into Messages column C. Input changes invalidate the saved snapshot. Approval resets after execution.
- Test failures and duplicate runs on disposable data before adapting the project.
No web deployment or public endpoint is needed for this bound-script project. There are no scheduled triggers or automatic retries.
Troubleshooting and expected outcomes
- Invalid JSON: use double quotes and remove trailing commas. Expect no updates until fixed.
- Missing/duplicate IDs: ensure each input ID appears exactly once in Response.
- Unknown category: copy an exact value from Categories. Capitalization matters.
- Changed input but unchanged answer: the fixture is fixed. Write a candidate yourself or configure the exported script.
- Storage unavailable: export a JSON backup. Saving is explicit, not automatic.
- Provider error: inspect account/model permissions and quotas. The script reports HTTP status and stops without writing classifications.
- Wrong classification despite passing tests: validation checks structure. Human review checks meaning.
- Approval rejected: regenerate or update the snapshot after changing the Messages rows, then review again.
Compatibility and privacy
Reference check: 17 November 2026. Models, access and quotas change. Configure a supported model rather than relying on a hardcoded model name. The website requires modern JavaScript modules and IndexedDB. Your data remains local unless you copy it or run the configured Apps Script export.
Gemini generateContent reference · Current models · Script Properties · Apps Script quotas · Gemini in the Apps Script editor
Gemini editor assistance is different from making a Gemini API request in your application. This export uses the latter. No live Google integration has been tested in this package.