Templates and reusable workflows

The real productivity gains come when great prompts become reusable assets your whole team can rely on.

5.1 From one-off prompts to reusable templates

If you do a task more than a few times, turn your best prompt into a template with placeholders for the parts that change. Teams can share templates so everyone gets consistent, high-quality output.

TEMPLATE: Customer reply

You are a friendly, professional customer service agent for {{company}},
a UK kitchenware brand. Our tone: warm, concise, never defensive.

Customer message:
<message>{{customer_message}}</message>

Relevant policy:
<policy>{{policy_extract}}</policy>

Write a reply that:
1. Acknowledges the issue in the first sentence
2. Explains what we'll do, based only on the policy
3. Gives a clear next step and timescale
Under 120 words. UK English. If the policy doesn't cover the issue,
write "ESCALATE" and a one-line reason instead of a reply.

5.2 Where to save your instructions

Most assistants now let you store standing instructions and context so you don't repeat them:

• Custom instructions / personalisation — your role, preferences, UK English, preferred formats.
• Projects (for example in ChatGPT and Claude) — shared files and instructions for a body of work.
• Custom assistants such as custom GPTs, Gemini Gems or Copilot agents — packaged instructions and knowledge for a specific task.
• Skills and agent instructions — reusable procedures some tools can load when relevant.

Check your organisation's rules before sharing custom assistants or uploading documents to them.

5.3 Treat important prompts like products: test them

For prompts that run often or matter a lot — especially inside business processes — "it looked fine once" isn't enough. Anthropic's guidance recommends defining success criteria and building tests before refining prompts.

1. Define success: what makes an output good? (accurate, under 120 words, correct category, polite tone, escalates when it should)
2. Build a test set: 10–30 realistic inputs, including tricky edge cases (angry customers, ambiguous requests, missing information).
3. Run and score: check each output against your criteria.
4. Change one thing at a time, re-run, and keep the version that scores best.
5. Re-test after model updates.

5.4 Version and document your prompts

Keep a simple prompt library: the prompt, what it's for, its owner, the date, the model it was tested on and known limitations. OpenAI's developer guidance recommends keeping production prompts in version control alongside application code, with tests. That way you know what changed when results change.

5.5 Prompts in automated systems

When a prompt runs without a person reading every output — in an automation, chatbot or agent — design for failure:

• Give an explicit way out ("ESCALATE", "UNSURE") and route those cases to a human
• Validate the output format before using it
• Treat any user-supplied or web content as data, not instructions — it may contain prompt injection
• Log inputs and outputs (in line with data protection rules) so you can investigate problems

Key takeaways

• Templates with placeholders make quality repeatable.
• Use custom instructions, Projects and custom assistants for standing context.
• Define success, build a test set, change one thing at a time, re-test after model updates.
• Keep a documented prompt library.
• In automation, design explicit escape routes and treat external content as data.