Prompting 103: context คือ prompt ที่แท้จริงPrompting 103: context is the real prompt

When AI gives a bland answer, the problem is often not the model.

The problem is that it does not know the work context.

Context is anything the model needs to answer correctly:

• source text
• audience
• goal
• background
• constraints
• examples
• definitions
• current facts
• what has already been tried

Context is not a data dump

More context is not always better.

Weak:

Here is a 40-page document. Summarize it.

Better:

Summarize the pasted policy for new employees.

Focus on:
- what they must do in their first week
- who they should contact
- rules that affect daily work

Ignore:
- legal boilerplate
- internal document history

Output:
- Thai summary
- 8 bullets maximum
- include page references if visible

The better prompt tells AI which parts of the context matter.

Use labels

Labels help the model understand boundaries.

TASK:
Rewrite the announcement.

AUDIENCE:
Thai office workers, age 25-40.

STYLE:
Clear, warm, not playful.

SOURCE TEXT:
[paste draft]

Markdown headings, bullets, and separators are enough for most everyday prompting.

Put source text in a container

When you paste source material, separate it from instructions.

Use only the source text below. If the answer is not in the source, write "not found".

SOURCE:
---
[paste document]
---

QUESTION:
What is the refund deadline?

This reduces accidental mixing between your instruction and the document.

Tell AI what not to use

Sometimes the most important context is a boundary.

Use only the pasted text. Do not use outside knowledge.

Use current web sources. Do not rely only on memory.

Do not invent names, numbers, dates, or policy details.

Different jobs need different boundaries. A brainstorming task can use imagination. A policy task needs source discipline.

Give current facts

Models may not know what changed today. For current information, either provide the facts yourself or ask the AI to search if the tool is available.

Weak:

What is the latest policy?

Better:

I need a current answer. Search official sources first. Cite the source and date. If sources disagree, say so.

If you already have the source:

Use this May 2026 policy update as the source of truth:
[paste source]

Practice: context diet

Take a long messy request and split the context into three groups:

Must include:

Helpful but optional:

Do not include:

Then write a prompt that includes only the first group and maybe one or two items from the second group.

References

• Anthropic: Effective Context Engineering for AI Agents frames modern AI work as managing the right information in the context window.
• OpenAI Prompt Engineering describes context as a way to supply data outside the model's training set and constrain responses to selected resources.
• Dong et al., 2025: Survey of Context Engineering surveys retrieval, processing, and management of context for LLM systems.