Issue 12: Issue 12: Prompt engineering, the small stuff that matters

What shipped

Anthropic publishes 2026 prompting practice guide for non-developers

Anthropic released an updated prompting guide aimed at non-developer users, covering role framing, structured outputs, and guardrail instructions for daily knowledge work.

Source →

OpenAI ships custom GPT improvements for organisation-wide deployment

OpenAI released organisation-wide custom GPT controls, allowing IT teams to publish approved prompt-and-knowledge bundles for staff use.

Source →

Google Workspace adds Gemini side-panel improvements

Google updated the Gemini side-panel in Workspace with persistent project memory, named source citations, and a one-click prompt history view.

Source →

Four prompting habits any professional can adopt this week.

This week is for the general professional. The four habits work across roles, sectors, and tools. Each is achievable in a single sitting and produces a noticeably better output the next time you use it.

One. Start every non-trivial prompt with a role statement. "You are a [role] supporting a [user] preparing [output]." Three lines that change the entire response.

Two. Name your inputs and number your outputs. The model performs noticeably better when it knows what it is reading and what it must produce.

Three. Include one guardrail per prompt. "Do not invent." "Where evidence is insufficient, state what would be needed." "If a name appears, stop." Specific, not generic.

Four. Save the prompts that work. Three folders. Drafting. Summarising. Analysing. Reuse beats reinvent every time.

These habits compound. Role and inputs improve quality. Guardrails reduce risk. A reusable library cuts setup time. Together they shift AI use from occasional novelty to repeatable practice.

Why structured prompts beat clever prompts for professional work.

Visual 1. Indicative quality lift from structured prompting components, cross-role survey, April 2026. Sources: Anthropic prompting guide, internal pilot summaries from three Australian employers. Indicative only.

Prompt of the week.

Setup: This is a meta-prompt. Use it to convert any one-line task description into a structured, reusable prompt with a role frame, named inputs, numbered outputs, and a guardrail. Paste the rough task. The model returns a fully formed prompt you can save.

You are a prompt engineering coach supporting a knowledge worker who wants to convert a rough task into a reusable, structured prompt. Inputs I will provide: - A one-line description of the task. - The role of the user (for example, HR generalist, claims officer, finance analyst). - The intended output type (document, list, table, summary). - Any sensitivity rules that apply (for example, no personal information, no client identifiers). Produce: 1. A fully formed prompt the user can paste into Claude, ChatGPT, or Gemini, with a role statement, named inputs, numbered outputs, and at least one guardrail instruction. 2. A two-line note explaining what to substitute when reusing the prompt. 3. A one-line success criterion for the prompt, written so the user can self-check whether the output is good enough to use. Do not invent inputs the user did not list. Where evidence is insufficient, ask one targeted question to clarify before producing the prompt. If the rough task involves regulated or sensitive data, include a sensitivity guardrail in the produced prompt by default.

How to use it: Paste a one-line task description. Add your role and the output type. Run the prompt. Save the result. Reuse it next week with a different input set. The library you build over a quarter is the most valuable AI investment most professionals make.

Risk: Models will sometimes pad the produced prompt with redundant filler. Edit ruthlessly. The goal is the smallest prompt that produces the output you need every time. If the produced prompt does not work twice in a row on slightly different inputs, it is not yet a reusable prompt.