If you frequently use AI for writing, summarizing, analyzing, or planning, you may have encountered the same issue: in the first chat, the results are quite good, but when the task continues in another conversation, AI starts asking basic questions again or produces answers that go in a different direction.
The problem is often not because the model suddenly becomes 'dumb'. The necessary context does not always carry over to the new conversation. A practical solution is to create a context pack: a short document containing basic information, goals, rules, examples, and work formats that can be attached to various prompts.
This concept aligns with the official guidelines on prompts that suggest users provide clear instructions, sufficient context, and then iteratively refine the prompts based on the results. OpenAI also suggests breaking complex tasks into smaller parts.
What is a context pack?
A context pack is a 'context package' for a specific type of work. Its contents are not long prompts asking AI to complete everything at once, but rather information that helps AI understand the situation before starting work.
Think of it as handing over a task to a new coworker. You wouldn't just say, 'Please write the article.' You might also explain who the audience is, the writing style to be used, terms to avoid, available materials, and what constitutes a good outcome.
A context pack plays the same role. The difference is that its format can be saved as notes and used repeatedly.
Most Useful Contents of a Context Pack
You don't need to create a dozen-page document. For routine work, one to two pages is usually sufficient. Use the following sections as a framework:
- Work Objective: explain what you want to achieve, not just the activity.
- Background: provide information that AI needs to know before starting work.
- Target Audience or Users: mention their level of knowledge, needs, and problems.
- Style and Limitations: for example, natural Indonesian language, not excessive, without unverified claims.
- Source Materials: list data, notes, links, or documents that can be used.
- Examples of Desired Outcomes: provide one or two examples so AI understands the pattern, rather than just reading abstract instructions.
- Things to Avoid: for example, jargon, repetition, specific formats, or unfounded assumptions.
Google's prompt guidelines also emphasize the importance of context, examples, consistent formats, and iterative processes. Good examples can help the model capture the style, structure, or level of detail you want. Google explains this strategy in its prompt design documentation.
Ready-to-Use Context Pack Template
Here is a simple template that you can copy into your notes application:
CONTEXT PACK: [Job Name]Main Objective:
[Desired outcome]
Background:
[Important information about the project, product, audience, or issue]
Target Audience/Users:
[Who they are and what they need]
Work Style or Language Style:
[Example: clear, warm, concise, not sensational]
Important Rules:
- [Rule 1]
- [Rule 2]
- [Rule 3]
Available Materials:
[Data, notes, documents, or links]
Examples of Preferred Outcomes:
[Brief example]
Things to Avoid:
[Errors, terms, assumptions, or unwanted formats]Note that this template does not ask AI to perform a specific task. It only prepares the working conditions. Specific tasks should be given after the context pack is attached.
Use Two Layers of Prompts
A neater workflow usually separates prompts into two layers. The first layer is relatively stable context. The second layer is the variable task.
For example, you have a context pack for managing blog content. Each time you want to create a new article, you don't need to repeat the editorial style, target audience, and structural rules. You just need to add a request like:
Use the context pack above to work on the following task.
Task:
Create a 900-word article outline on how to choose a financial tracking app for families.
Before writing, display:
1. The assumptions you are using.
2. Information that is still lacking.
3. The proposed structure of the article.The request to 'display assumptions and information that is still lacking' is useful for reducing rework. AI is given the opportunity to highlight unclear parts before generating a lengthy draft.
Do Not Include Everything in One Pack
A context pack that is too long can also be problematic. Irrelevant information makes the main instructions hard to see and increases the likelihood of AI mixing rules from multiple projects.
Create several packs based on the work, rather than one giant document for all needs. For example:
- Writing context pack: language style, audience, structure, and editorial rules.
- Analysis context pack: metric definitions, data sources, assumptions, and how to present findings.
- Coding context pack: project structure, language version, naming standards, and dependency limitations.
If a piece of information only applies to one task, do not include it in the permanent pack. Place it as an additional note on that prompt.
Add Examples, But Don't Overdo It
Examples help AI understand intentions that can sometimes be difficult to explain with rules. For instance, the sentence 'write in a non-stiff manner' is still quite subjective. An opening example that you like will provide more concrete guidance.
However, too many examples can lead to results that mimic the surface of writing without understanding the purpose. Choose examples that represent important patterns: paragraph length, word choice, how to explain terms, or how to structure recommendations.
Use examples from your own work, and ensure that you do not include personal data, credentials, customer information, or internal documents that should not be shared with AI services.
Context Pack Is Not a Guarantee of Always Correct Answers
A context pack makes work more consistent, but it does not turn AI into an automatic source of truth. The model can still misunderstand data, draw unsupported conclusions, or convey information that sounds convincing but is incorrect.
OpenAI reminds us that ChatGPT can generate incorrect or misleading information with a confident tone. Therefore, for topics related to law, health, finance, security, or current information, results still need to be verified against reliable sources.
What You Can Do Now
- Choose one task that you repeat most often with AI.
- Create a context pack of no more than one page using the template above.
- Save the first version in a notes application or document that is easy to find.
- Test it with two or three different tasks.
- Note the parts that often lead AI astray, then refine the context pack.
The ultimate goal is not to create the longest prompt possible. The goal is to transfer important context from your head into a reusable format. This way, each new conversation does not start from a blank page, and AI can work with a more consistent understanding.
Sources & Further Reading
– Rio Yotto @rioyotto