Working backwards
from the outcome.
Prompt engineering for people who need it to actually work — today, in your Pod.
/upskillinglabs0516
AI can produce anything.
Prompt engineering is how
you stop it from doing that.
Same AI. Same minute.
Help me with my recycling project.
5 ideas for a DC "can I recycle this?" tool. One sentence each. Works in under 3 taps. No app download. For a renter, not a homeowner.
The AI isn't broken. The spec is.
What is a prompt?
Anything you give the AI so it can do the work.
A question. A task. A doc. A transcript. An example. Usually a few of those, in sequence.
Probability machine.
No defaults.
A search engine returns what exists. AI returns what's probable, given what you fed it.
Prompt engineering = shaping what's probable until it matches what's useful.
It's design thinking.
You already know how.
If the output fails, don't blame the AI.
Go back to Define.
Stop asking:
"How do I phrase this?"
Start asking:
"What does done look like?"
OSCAR.
Five pieces of a spec. Fill them before you type.
Outcome.
Describe the artifact, not the task.
"Help us with digital carbon."
"A 5-lesson curriculum + a clickable map of data centers in our community."
"Help with rebates."
"A four-stop journey — awareness, decision, funds, trust — for a middle-income homeowner."
"Help with recycling."
"Three screens: photo→answer, habit-swap nudges, proof of what actually got recycled."
Success.
3 checks. Your grading rubric.
- ✓Each lesson reads in ≤ 5 min
- ✓Map zooms to neighborhood level
- ✓Curriculum and map point at each other
- ✓Names rebates eligible for THIS household
- ✓Shows when the money actually arrives
- ✓Surfaces 2 trust signals per vendor
- ✓Photo → answer in ≤ 3 taps
- ✓Names cheapest sustainable swap
- ✓Cites a real "got recycled" example
Constraints.
What it must not do.
Curious citizen, not climate-doom.
5 lessons. Phone-readable.
Invent kWh figures or center coords.
Neighbor, not salesperson.
One screen per stop. Four stops.
Cite dollar figures it can't source.
Direct, never lecturing.
Yes/no first, detail after.
Punt to "check with your hauler."
Assumptions.
Name what you're letting it guess.
Assume Reader has never heard "data center."
Don't Guess my metro — I'll paste it.
Assume Reader owns or co-decides on a home.
Don't Guess their state — I'll paste the ZIP.
Assume Reader is in a kitchen, skeptical, deciding now.
Don't Guess the city — I'll specify.
Reference.
Hand it the source. Don't make it guess.
- IEA AI-energy report
- EPA data-center efficiency data
- Local news on data-center siting
- DSIRE incentive database
- State PUC + rebate pages
- ENERGY STAR fact sheets
- City solid-waste rules
- Montgomery Recycle public reports
- EPA recycling guide
If it hallucinates, R was empty. Almost every time.
You are not prompting a model.
You are writing a one-page spec
for an infinitely fast,
probabilistic intern.
OSCAR a prompt you'd run this week.
- 1 Pick a real task. Or use the worksheet.
- 2 Fill the card. 4 minutes.
- 3 Run it. Compare to a zero-spec version.
Four Pods. Four small interventions.
One framework.
Grab the master prompt at /upskillinglabs0516/prompt. Pick the card closest to your Pod. Brain-dump. Iterate.
When one prompt isn't enough.
Same hour of work. Exponentially more reach.
One ask before you leave today.
Pick the smallest version of your Pod's problem. OSCAR it. Run it.
Brain-dump goes in. OSCAR plan comes out. Iterate from there.