Context Is a Resource You Control
A working note on the thing that shapes an answer more than any phrasing: what you put in front of the model.
Students spend enormous effort on how to word a prompt and almost none on what to put in it. That's backwards. What you place in the context (the document, the examples, the background, the prior turns) shapes the answer more than any clever phrasing, and unlike the model's training, the context is the part you fully control.
The common move
All the effort goes into the instruction ("really good") and none into what the model would need to do it: the posting, your background, the work you're proud of. The question worth asking: what does the model actually have to work with?
Wording the request, versus supplying the context
- Polishing the instruction. Reworking "make it good" into "make it compelling and professional."
- Assuming the model knows you. Expecting a fit to facts it was never given.
- Blaming the phrasing. When it's generic, you reach for better adjectives.
- Giving it the raw material. The posting, your resume, the two projects that fit.
- Filling what it can't know. Your specifics are not in its training; only you can add them.
- Reading generic as starved. A vague answer usually means thin context, not weak wording.
A generic answer is most often a context problem wearing a phrasing costume. The model isn't being lazy; it's working with what little you handed it.
The more honest move: feed it before you instruct it
Before polishing the request, ask what raw material the task actually needs, and put that in front of the model. Context first, instruction second.
The "using only what I've given you" instruction matters: it ties the output to your material and turns invented specifics into visible brackets instead of confident fabrications (which connects to the verification habit from sheet 4). You did less wording work and got a far better result.
Try this
Before improving your phrasing, ask: what would a capable person need to do this task that I haven't provided?
Put that material in the prompt first. Then make the request.
If the answer is generic, suspect thin context before you suspect weak wording.
There's a subtler effect, too: the examples and material you supply quietly steer what kind of answer you get. A later sheet takes that up.
The principle underneath
What you put in the context shapes the answer more than how you word the request, and it's the part you fully control. Feed the model before you instruct it.