The skill that beat prompts
Same request the whole session. We only change what the model can see.
Aki Wijesundara, PhD
Instructor · The AI Internship
We will approve or deny a refund without rewriting the request. Everything happens in Context Lab.
Instructions, knowledge, and state
Add the right documents — then break it with clash
Memory helps. Noise distracts.
One toggle that proves context is measurable
Context engineering is all three — plus a budget.
How you word the task. Role, tone, rules, few-shots. Everyone's 2023 obsession.
What the model can look up. Docs, policies, records, retrieved facts.
What it remembers. History, memory, prior turns, scratch notes.
The job stopped being writing and became assembling.
Should we approve a refund for order #48291? The customer says the item arrived defective.
Press Run. Watch what changes when the request does not.
Context Lab · Shift + ← / → for slides
You changed what the model could see.
Current policy + order + prior tickets → approve, cite the defective-item exception, note 44 days and two tickets.
Add the 2024 archive and two documents disagree. Nothing says which wins. Dumping everything fails in production.
A conflict-resolution rule in Instructions. Clean answer again — and it names which policy it applied. Retrieval quality and conflict handling beat prompt polish.
State can help — or steal every token that mattered.
A retention note shifts tone and explains why. No prompt change. No new document.
2,400 tokens of unrelated history. Answers get vague — or mimic the transcript instead of deciding.
Summarise old turns. Keep the last few verbatim.
Give the decision its own clean sub-agent.
Write scratch state to a file. Read back only what's needed.
Context is not a text box. It is a system with inputs you can tune and measure.
Policy + order + tickets + conflict rule → strong, citable decision.
The defect is no longer corroborated. Same prompt. Lower confidence.
Measure context quality the same way you measure any other system input.
When the answer is wrong, ask what it could not see — before you ask how to reword it.
More context is not better context. Clash, staleness, and noise cost you accuracy.
Assemble on purpose: instructions, knowledge, state — each with a budget.
The Agentic AI Builder's Bootcamp is building the systems that make these decisions all day — with evals to prove they work.
maven.com/tailabs/ai-builder
Aki Wijesundara, PhD · The AI Internship
RAG is one column — knowledge. Context engineering is the whole window: instructions, knowledge, and state, under a budget.
No. Distraction and clash get worse with more room, not better. Attention is still finite.
You measure it. That number is an eval — and that is the next skill to learn.