When I first wrote about Context Engineering in 2023 (I am starting to think I might have been the person to coin that term since I can’t find any earlier mention of it), I created some key concepts that I’ve used ever since.
Today we are covering Context Sensitivity
At the core: how much novel context do you need to provide to an LLM to get it to generate an output that aligns with your desired goal
Understand it in a simple example:
“what is 2 + 2?” is a low context sensitivity prompt. Lets make context sensitivity into a simple formula:
(Model : Scenario) * Context() → Output
So if we use our 2 + 2 example:
Ask GPT-5 “what is 2+2?”
(GPT-5 : Arithmetic[what is 2+2?]) * Context(Nothing)Will get you the same you the same as:
(GPT-5 : Arithmetic[what is 2+2?]) * Context(“Here is how you add numbers together: 1+1 = 2”)
You do not need to insert any novel context to get a result that aligns with your goal
(unless it is somehow other than getting it to generate 4 lol)
A low sensitivity prompt is bounded. It has a clearly defined problem, desired outcome and makes few assumptions and no layered assumptions
So what is a high context sensitivity query?
“Across all my last 30 call onboarding call transcripts, what were the top 5 major trends?”
It may seem like all you need to do is provide the last 30 calls. Nope.
First hidden variable: what sort of trends? What are you measuring to then map into a trend?
Total call length? That is pretty useless. Ok so maybe pain points mentioned?
Second hidden variable: Ok so how do we group pain points? Do we have a short list of types already, or do we really not know and we want the model to hypotheses and validate?
Third hidden variable: How should it go about validating these hypotheses? What criteria?
Should it try to search for key words across all the full transcripts? What about dumping them into a single context window.
Well, you really can’t get high quality results just dumping 30 call transcripts into a single context window, so you really need to:
define what you do, why you do it and how this informs what you measure
define how and why you measure so you can know what a trend looks like
create a one agent to one transcript framework using the above
use an agent/context window to summarize, classify each transcript, ensure it has a standardized output schema
then take all of these classification + summarization objects and feed them into an orchestrator agent
create a framework for the orchestrator agent to do something with these compressed versions
(you should probably tell it to review, hypothesize a handful of trends and then spawn a subagent to go traverse through the full transcripts to validate those hypotheses. An agent for each hypothesis.)
then your hypotheses subagents report back with findings to the orchestrator agent
And then your orchestrator agent has all 30 summarized + classified + validated hypotheses, so finally you can have it write up a validated, grounded, and actually usefully actionable trend report.
That’s what a high context problem feels like. You peel back the layers like an onion and more emerge as you wrestle with the problem set.
A high context sensitivity prompt is unbounded: There is no built-in, well defined stopping point to find a desired outcome and the prompt forces the LLM to make assumptions to create a desired outcome, or you have to build a boundary around the problem.
You could keep slicing and dicing into smaller and smaller hidden variables, but at a certain point it becomes counter-productive and pedantic. This is part of the craft. You have to figure out where to set the bounds of the problem.
It is both ruthlessly objective via trial and error and completely based on intuition that you can figure out the right approach for the right context sensitive problem.