Give Your Agents the Right Context

Difficulty: Advanced · Estimated time: 20 min

What You'll Learn

How to build a context pack — the background material, examples and glossary that stop an agent's output sounding generic — and how to keep it from going stale.

Why This Matters

Solid instructions tell an agent how to behave. They don't tell it what's true about your world — your organisation's terminology, what a good example of its output actually looks like, or the facts it should never get wrong. Without that, even a well-structured agent produces plausible-sounding generic work instead of output that fits your actual situation.

The Concept

A context pack is the second half of a good agent, sitting alongside the instructions from Guide 17. It typically has four parts:

Background facts — things that are always true and the agent should never have to guess (your team structure, your product names, standard pricing, your client's industry).

Examples of good output — two or three real, approved examples of the agent's exact job done well. These teach tone and shape far faster than a description can.

Constraints and style notes — house style, things to avoid, legal or brand lines it must respect.

A glossary — terms specific to your world that a generic model wouldn't know (internal product codenames, client shorthand, acronyms your team uses).

This lives wherever the instructions live: as uploaded files in a Project's knowledge, or bundled as resources inside a Skill.

Step-by-Step

List the background facts the agent would otherwise have to guess at or ask you for every time.

Collect two or three gold-standard examples of the finished output — real past work you'd happily show someone as "do it like this."

Write a short style and constraints note — anything you'd correct a new hire on in their first week.

Build a glossary of terms, acronyms and names specific to your situation.

Bundle it into the same Project or Skill as the instructions from Guide 17.

Set a review point — a date or trigger for checking the pack is still accurate.

Example

For a client-facing proposal-drafting agent at a marketing agency:

Background facts: the agency's three service tiers and what's included in each; standard turnaround times; the client's industry.

Examples: two previously approved proposals, with a one-line note on why each worked.

Style notes: no jargon in client-facing sections; always lead with outcomes, not deliverables; agency name always spelled out in full on first mention.

Glossary: internal shorthand like "the audit" (meaning the discovery call) and "Tier 2" (meaning the mid-tier package), so the agent uses them correctly rather than guessing.

Try This

For the agent you've been building since Guide 16, list three background facts it currently has to guess at, and find one real example of the output done well that you could hand it.

Copy This

CONTEXT PACK FOR: [AGENT NAME]

BACKGROUND FACTS:
- [FACT THE AGENT SHOULD ALWAYS KNOW]
- [FACT THE AGENT SHOULD ALWAYS KNOW]

EXAMPLES OF GOOD OUTPUT:
1. [LINK OR PASTE EXAMPLE 1] — works because [REASON]
2. [LINK OR PASTE EXAMPLE 2] — works because [REASON]

STYLE AND CONSTRAINTS:
- [HOUSE STYLE RULE]
- [THING TO AVOID]

GLOSSARY:
- [TERM]: [WHAT IT MEANS IN THIS CONTEXT]
- [TERM]: [WHAT IT MEANS IN THIS CONTEXT]

LAST REVIEWED: [DATE]
NEXT REVIEW DUE: [DATE OR TRIGGER, e.g. "when pricing changes"]

Common Mistakes

Dumping an entire company wiki in as "context" instead of the facts actually relevant to this one job.

Using mediocre or outdated examples because they were the easiest to find.

Building the pack once and never updating it after prices, names or processes change.

Treating the glossary as optional — it's often what makes output sound like it was written by someone on the inside.

Level-Up Tip

Date every context pack and keep a one-line changelog at the bottom of it. When output quality suddenly shifts, the changelog is the first place to look.

Action Checklist

○Background facts listed

○Two or three real gold-standard examples collected

○Style and constraints notes written

○Glossary of situation-specific terms built

○Review date or trigger set

Next Guide

19 — Build an Entire AI Department, where single agents become a coordinated team.

Premium Content

Levels 2–6 require a mygyaan membership to access the full guides.

₹1,000/month — full mygyaan.ai access (all modules)

₹3,000 one-time — this course only

What you will learn