Working Notes on Prompting
introductory series, version 1.2. Subtitle: What the Tool Is Actually Doing. Ten sheets for the creative coding and AI research course.
Open this first to orient. Each sheet addresses one common prompting belief or habit, shows what it actually does, and offers a more honest move plus a principle that generalizes. Click any sheet title to open it. Readable in five minutes each.
How to read the series
- Foundation (sheets 1 to 5). The model is not a person, your intuitions mislead you, and you are responsible for what you do with its output. This group also teaches the two active skills the rest assume: verifying, and iterating.
- Method (sheets 6 to 7). How you actually supply the model: show examples rather than describe, and feed it context before you instruct it.
- Applied (sheets 8 to 10). The orientation habit applied to specific technique families: personas, length, and step-by-step reasoning.
- Sheet 10 closes the set with a short reflection gathering the thread through all the sheets. If the series grows, that closing moves to the new last sheet.
- The through-line: every sheet makes the same move in a different costume. Stop asking what words to type; ask what the model is actually doing, and stay responsible for checking it.
Foundation: the model is not a person, and you are responsible
The "you are an expert in [domain] with [N] years" framing. What it does (shifts vocabulary and register) versus what it doesn't (add knowledge or reliability), and the more honest peer-engagement move when you already know the domain.
Whether saying "please" does anything. No special mechanism, but polite requests carry the clarity and specificity that actually help, so it correlates with better answers without causing them. Aim at the lever, which is specificity.
Why telling a model not to make things up doesn't work: hallucination is a property of how it generates text, not a choice. Includes a real confabulation case from this course's development, a grounding-and-flagging prompt, and the case for a verification step.
How to actually check an output, since "just verify it" is harder than it sounds. Triage by what's load-bearing, seek independence rather than agreement, try to disconfirm. Includes using a second model as a skeptic, not a rubber stamp. Deepens sheet 3.
The habit the other sheets assume: read what came back, diagnose why, change the cause rather than the symptom. A weak output is evidence about your prompt, not a result to discard and re-roll. The foundational skill most advice skips.
Method: how you actually supply the model
Why describing the output you want (stacking adjectives) works worse than showing it. The model resolves abstractions by averaging; a few concrete examples (few-shot prompting) give it a pattern to match instead.
What you put in the window shapes the answer more than any phrasing, and unlike the model's training, it's the part you fully control. A generic answer is usually a context problem in a phrasing costume. Feed the model before you instruct it.
Applied: technique families
Generalizes sheet 1 to the whole persona family ("act as," "you are," "explain like I'm 10"). A persona shifts register, not capability, and "explain like I'm 10" often buys condescension. Name the behavior you want, not the character.
What "be concise" and "in 500 words" actually do. The model approximates length and routinely misses, and "concise" silently strips substance. Separate content from length: generate the substance first, then edit down.
The one technique that genuinely works, for some tasks. Step-by-step is working room, not added intelligence: it helps when a task has steps worth showing and is noise when it doesn't. Carries the movable series-closing reflection.
How to use this series
The sheets share one structure: name the common move, compare what you think it does against what it actually does, offer a more honest move with a usable prompt, and close on a principle that generalizes past the specific technique. The comparison block is the centerpiece of each sheet, with the misconception on the left and the correction on the right.
The series is undergraduate-first, but the principles travel. Examples lean lightly on design-studio context; the moves apply to any field. Sheets can circulate to a more general audience with little or no change.
Sheet 1 predates the others and was elevated to the current design system (v2.1). Its comparison is a "what it does / what it doesn't do" split rather than a misconception/correction split, so its column treatment differs slightly by design.
This is the introductory tier. A separate intermediate tier ("Working With What You Can't Trust") is planned, going deeper rather than rehashing: how the model can't reliably report its own reasoning, how to prompt for calibration over confidence, and why fluency can beat truth. See the series log for status.
The set is built to expand. New sheets join the appropriate group as the course surfaces new techniques. When a sheet is added, the series mark ("of ten") updates across all sheets, and the closing reflection on the last sheet moves to whatever becomes the new last sheet.