Prompt Engineering Guide for Daily Work

Purpose

Provide operational, copy/paste prompting patterns for daily work across major assistants, with explicit evidence boundaries and fail-closed behavior.

Use this page in daily-work AI workflows when:

You see confident outputs that aren’t supported by evidence (need an explicit evidence boundary + fail-closed behavior).
You’re working with long inputs (docs/threads/code) and need explicit handling for context limits and skipped content.
Tool behavior changes by runtime (browsing/connectors/code execution) and you need explicit tool-policy constraints.
You want to reduce agreement bias (sycophancy) and enforce “conflict with evidence → say so + cite”.

Choose the right tool (vendor-documented orientation)

Use the mapping below as a default decision guide. Verify availability by plan/tier in the vendor documentation.

ChatGPT (OpenAI): general-purpose drafting, summarizing, brainstorming; do not treat output as a sole source of truth.
Gemini in Google Workspace: when your work artifacts are in Workspace apps (Docs/Sheets/Drive/Gmail/Meet/etc.) and you need in-context help.
Microsoft 365 Copilot: when your primary workflow is inside Microsoft 365 apps (Word/Excel/PowerPoint/Outlook/Teams/etc.).
Claude (Anthropic): when you need long-context reading/writing; context window depends on model/tier.
Perplexity: when you need web search + citations as the default interaction pattern.

Core assumptions (operational)

1) Output is not guaranteed to be correct by default

Treat outputs as generated text that requires evaluation and (when needed) external verification.

2) Context is finite

Long inputs may exceed the model’s context window; do not assume “full deep read” happened unless enforced procedurally.

3) Sycophancy exists

If the user strongly asserts a belief, some assistants may bias toward agreeing rather than challenging; use objective-mode prompting to reduce this.

4) Tool use is product- and policy-dependent

Web browsing, code execution, and document analyzers depend on the specific product/runtime policy and may require explicit instruction.

5) Clear goal + constraints reduce generic answers

Define goal, audience, format, and constraints; request explicit uncertainty handling.

Copy/paste prompt templates

Use the templates page:

Prompt Engineering — Daily Work templates

Minimal security note (daily work)

If you summarize or “reason over” untrusted documents/emails/chats, treat embedded instructions as untrusted content. Do not let document text override your instruction hierarchy.