Prompting as engineering
Take this if you have a prompt you rely on and no way of telling whether your last edit improved it.
What you’ll be able to do
You can build an evaluation set for a prompt, measure a change against it rather than against your impression, and detect a regression you would otherwise have shipped.
What it assumes
Regular use of an AI assistant, and enough Python to run a loop over a list of test cases.
- Price
- ₹299
- Length
- 2 weeks
- Sessions
- 6
- Effort
- ~4 h/week · 8 h total
- For
- Class 11–12 · College · Anyone
Adds it to My courses on your dashboard so you can find it again. Payment is not open yet — nothing is charged.
What’s taught, in order
- 01
The problem with trying it and seeing
Model outputs vary between runs, so a single comparison tells you almost nothing. Establishing this experimentally in the first session is what motivates everything after it.
- 02
Writing the test set first
Ten to twenty cases with what a good answer must contain, written before you touch the prompt. This is the step that converts an opinion about a prompt into a measurement.
- 03
Graders that are not another opinion
Deterministic checks — substring, format, schema, numeric tolerance — and where you genuinely need a model to judge, plus why that judge needs its own test set.
- 04
Change one thing, measure, repeat
Running the suite against prompt variants and finding the change that helped on average while quietly breaking two cases. Regression is the discovery this sprint exists to hand you.
What you build
Take a prompt you actually use, build an evaluation set of at least fifteen cases with pass criteria, then make three deliberate edits and score each one. Report which edit won, which case each edit broke, and the wording change that helped on average but caused a specific regression.
Check it yourself against this
- At least fifteen cases with pass criteria written before any prompt edits.
- At least half your graders are deterministic rather than a model's judgement.
- Three prompt variants scored on the same suite, with the numbers shown.
- You identified at least one regression that the average score hid.
- You can state one thing your suite cannot measure about the output's quality.
Then answer this
Your best-scoring prompt broke two cases the old one passed — how should you decide whether to ship it?
Two or three sentences, in your own words. If you cannot, the course is not finished — go back to the module it came from.
What this course does not do
This is evaluation discipline, not a prompt library and not a course on model internals. It also stops short of production concerns — cost tracking, latency, caching, monitoring a live prompt — which belong with deployment rather than with a first evaluation habit.
Where this leads
The free Prompt Patterns Library is the companion to this sprint: it gives you nineteen reusable patterns to start from, and this sprint gives you the method for telling whether your edits to them are actually helping.
FAQ
- Is this just prompt engineering with extra steps?
- It is prompt engineering with a measurement, which turns out to be the difference between improving and drifting. Most prompt advice cannot be checked; everything in this sprint is checked against a suite you wrote first.
- Which model does this use?
- The method is model-agnostic and the sprint deliberately has you run the same suite against two different models, because a prompt tuned to one model's quirks often scores worse on another — which is itself one of the more useful things to discover early.
- Do I need to pay for API access?
- No. The exercises are sized to fit within free tiers, and where a paid key would help, there is a manual path that uses the chat interface with a smaller test set. The habit transfers either way.
Other courses
This is a standalone course. It does not add weeks to the 78-week curriculum or change your roadmap. If you want the whole path instead, the full programme is ₹399/month · ₹2,000/year — see what it covers.