Foundations · Lesson 1
Give it what it cannot know
Most disappointing responses are not a reasoning failure. They are an information failure. You asked someone to do a job while withholding half the brief, and they did a reasonable job on the brief they got.
The naive attempt
Write a follow-up email to a client who hasn't responded.
This produces a competent, generic, slightly obsequious email. It is unusable, and it is unusable for a reason you can name: nothing in that request identifies which client, what they went quiet about, how long it's been, whether they owe you money or you owe them a deliverable, or whether the relationship is warm enough to be direct.
The model filled those gaps with the statistical average of every follow-up email ever written. That average is bland because the world's follow-up emails are bland.
What "context" actually means
Not "more words." Three specific categories the model has no way to obtain:
Situation. The facts of your particular case. Who, what happened, when, what's at stake, what came before.
Constraints. The boundaries you are operating inside. Length, tone, things you must not say, policies, formats, the fact that legal has to sign off.
Intent. What you want to be true after the reader reads it. Not the artifact, the outcome. "An email" is an artifact. "They reply by Friday with a date, without feeling chased" is an intent.
Most people supply some situation, no constraints, and no intent at all. Intent does the most work of the three, and it is the one most often left out.
The better version
Write a follow-up email to a client.
Situation: We delivered a website redesign three weeks ago. They approved the designs enthusiastically, then went silent when we sent the final invoice ($12k). Two emails unanswered. Our main contact, Priya, was promoted mid-project and may have handed this off. We've worked with them for two years and want to keep them.
Constraints: Under 120 words. No passive-aggression, no "just circling back." Do not threaten late fees. That's a decision above my pay grade. Do not apologise for following up; we did the work.
Intent: Priya replies with either a payment date or the name of whoever owns this now. She should feel this is easy to answer, not that she's in trouble.
The second version isn't longer because verbosity is good. It's longer because each addition removes a specific way the output could go wrong. "Don't say just circling back" exists because that phrase was nearly certain to appear. "No late fee threat" exists because a model told the invoice is unpaid will reach for escalation.
The counter-skill: what to leave out
Over-contextualising is a real failure, and it looks like this:
You are an expert professional communicator with 20 years of experience in client relations. Email is a written form of communication that is important in business. It is important to be clear and professional, and to consider your audience carefully.
None of that is information the model lacks. It's filler that dilutes the parts that matter, and on longer prompts it actively competes for attention with your real constraints.
The test for any sentence in your prompt: could a competent stranger have written this sentence without knowing anything about my situation? If yes, delete it. It's teaching the model something it already knows better than you do.
Where this doesn't help
Context cannot fix a task the model genuinely cannot do. If you need this quarter's actual revenue figure, no amount of framing produces it. You need to supply the number or give the model a way to look it up. Adding context to a task that needs data is a common and frustrating loop: the response gets more confident and no more correct.
Distinguish "it doesn't know my situation" (fixable with context) from "it doesn't have this fact" (fixable only by supplying the fact).
The move
Before you send anything non-trivial, write three lines:
- Situation: what's true that it can't infer
- Constraints: what would make the output wrong even if it's good
- Intent: what should be true after someone reads this
Ninety seconds. It is the highest return habit in this entire course.
Exercise
You want help rewriting a rejection email to a job candidate. Write the prompt you would actually send, including the context the model needs and cannot work out for itself. Write the prompt, not the email.
How this gets marked
- 25%Says who the reader is and what they already know about the situation.
- 25%Says what the real constraints are. Tone, length, what must not be said, policy limits.
- 25%Says what should be true after the reader reads it, not just what artifact to produce.
- 25%Does not waste words explaining things any competent writer already knows.