Prompt engineering

Prompt engineering is simply me telling an AI exactly what I want, in the order and language it understands best. I picture the model as an eager but literal intern; if I hand it vague direction, it returns vague work. So I write prompts the way a film director calls a scene: I set the backdrop (“You’re a growth strategist”), define the action (“Draft a LinkedIn hook for CFOs”), sprinkle style cues (“British English, 12 words max”), then hand over the megaphone. When those three puzzle pieces—scene, job and success criteria—click together, the model stops guessing and starts performing.

Great prompts unlock five compounding advantages. First, predictable quality: when my instructions are tight, I skip the roulette of endless “Regenerate” clicks and get usable copy on the first pass. Second, speed: clarity up-front means the model spends tokens creating rather than interpreting, so tasks that once devoured an afternoon now finish before the kettle boils. Third, leverage: the less time I pour into drafting outreach emails or reworking code snippets, the more head-space I win for strategic work. Fourth, competitive edge: most people still treat AI like a toy; by standardising my prompts I crank out on-brand assets at a rhythm freelancers can’t match. Finally, team clarity: a documented prompt is a living SOP—anyone in my orbit can reproduce the same result without psychic readings or Slack back-and-forth.

Ignore prompt craft and each benefit flips: quality wobbles, deadlines drift, trust in the tool erodes and revisions balloon once stakeholders spot mismatches late in the process. Bad prompts are silent saboteurs; they hide problems until re-writes are painful. I’d rather front-load ten minutes of thinking than sink an hour chasing fixes downstream.

‍

I start every session by writing the outcome in one sentence: “When this is perfect, I’ll have a 200-word product description that converts SaaS founders.” That line becomes my North Star and keeps the brief honest. From there I layer structure before style—role and task on top, context and constraints in the middle, then formatting cues at the bottom—so the model tackles substance first and polish later.

Next, I feed it examples. A single stellar paragraph or finished table row anchors tone and format far better than adjectives like “engaging” or “punchy”. Whenever I paste long data dumps or drafts, I wrap them in triple quotes to fence off each section; clear delimiters stop the model from remixing my raw notes into the final answer.Iteration happens in public. I store every prompt version in Notion alongside a one-line note on why it worked or flopped. Patterns emerge quickly: maybe the model drifts when I forget success criteria, or gets verbose when I skip a word-limit line. High-performing prompts graduate into automations—Zapier flows that draft weekly scorecards, personalise cold outreach, or summarise customer interviews while I sleep.

Finally, I grade every output against that original success sentence. If it passes, the prompt joins the library untouched. If it fails, I change one slice—scene, job or success—then test again. By tweaking in isolation I see cause and effect instead of stirring the soup. Over time the loop tightens: clearer thinking yields sharper prompts which, in turn, sharpen my thinking. That fly-wheel is why prompt engineering isn’t a trick; it’s a habit that compounds like interest.