Most prompt libraries are collections of useful wording. That can help, but it is not enough when the work touches invoices, leases, payroll, insurance, access, or operating reports. In those settings, a polished wrong answer is still wrong, and it can cause a long week.

The Daft Gemini Playbook contains 112 Google Gemini prompts for Workspace built around one rule: prompts prepare; people decide. The prompts can compare lease terms, flag invoice-coding questions, turn a report into an action list, or draft a message for review. They do not tell Gemini to send, delete, archive, pay, post, approve, grant access, or communicate externally. That boundary keeps a bad answer in the review stage instead of turning it into an executed mistake.

Each prompt is more than a paragraph to paste into a chat. It identifies the Workspace surface, owning role, required source material, risk label, human checks, and a follow-up prompt to refine the result. Yellow prompts require the owner to check sources, calculations, and assumptions. Red prompts add an authorized reviewer before an output is used externally or drives an action.

Reusable instructions also ask Gemini to cite the supporting file, tab, row, clause, or date for material conclusions. If the information is missing, it should say so. A neat answer without evidence may still be useful as a draft; it is not something to rely on.

The playbook is designed around a common failure mode: a good prompt is run without useful context, the model produces a fluent response, and confidence gets mistaken for evidence. Gemini can reason, organize, and draft. The person using it still needs to provide the invoice, lease, report, approval matrix, email thread, or spreadsheet tab that makes the work real.

Attaching those sources changes the ask from “tell me what to do” to “show me the exceptions, missing evidence, assumptions, and next owner.” The initial response is still a draft. The reviewer opens the cited clause and checks the math. Asking the model to double-check itself is not verification, even if it uses a reassuring tone about it.

There is a deliberate maturity path. A prompt becomes a saved Gem only after the work is stable, repeated, and consistently produces the same output format. A Gem becomes an automated Workspace Studio flow only when the trigger is dependable, the outcome is reversible, and human approval still fits in the process.

The question is not simply whether Gemini can do a task. It is whether the inputs are clear, the output can be checked, and the result can be unwound if it is wrong. If not, it stays a prompt. Automating an unclear process just moves the confusion faster.

The library lives in Markdown and Git so it can have history, review, and controlled changes. Every prompt has a permanent ID. The catalog is generated rather than hand-maintained, and a validator checks the prompt structure, risk labels, required sections, links, review dates, and obvious unsanitized data. Its test fixtures make sure those checks actually fail when they should.

The repository contains no real company data: no organization names, account numbers, leases, vendors, or production examples. It uses placeholders and synthetic examples instead. That makes the library safer to share and easier to adapt to the way a real team works.

The cost is some friction, and it is intentional. Naming sources, owners, and review dates takes longer than pasting a clever paragraph into a chat window. It also gives the next person a way to understand what happened and check the result.

The 112 prompts are a catalog, not a rollout plan. The repository identifies eight low-risk, high-frequency prompts as the place to start. Run those with their real owners for a couple of weekly cycles before opening the rest. The project is MIT licensed, so adapt the thresholds, policies, source names, and approval matrix to your environment. Keep the central idea intact: let the prompt do the preparation and leave the decision with the person who owns it.