Case study
Strong vs weak prompts
I built two portfolio sites with Claude. One came from short, open-ended prompts, the other from detailed briefs. Same person, same kind of project, very different instructions. Here are the actual prompts, quoted word for word, and what each one led to.
Compare the two designs side by side →The two sites
Stronger prompts
This site
- Session
- This portfolio's build session
- Period
- 25 – 26 Oct 2026
The numbers
Counted from both session transcripts with the same script, as of 26 Oct 2026.
| Measure | Weaker prompts | Stronger prompts |
|---|---|---|
| Prompts writtenExcludes background-task notifications, context summaries, and a re-sent duplicate. | 42 | 10 |
| Calendar span | 44 days | 2 days |
| Words in the opening prompt | 39 | 2,384 |
| Prompts that asked for tests or acceptance checksKeyword match on verify / validate / test / acceptance / definition of done, then checked by hand (one false match, a video filename, removed). | 0 | 4 |
Six things a strong prompt does
- 01Concrete goalDoes it say what to build, not just how it should feel?
- 02ContextDoes it give the audience, background, and source material?
- 03ConstraintsDoes it say what must and must not happen?
- 04Acceptance criteriaDoes it define what done looks like?
- 05VerificationDoes it ask for the result to be tested or checked?
- 06ScopeDoes it say what to leave alone?
Side by side
Round 01
Starting the project
Kick off a portfolio site from scratch.
Weaker prompts · 25 Aug 2026
Ok, I want you to really have this project outstanding and professional with very creative design. Please first of all have a README.md completed. Put it any related works etc. Then next work on the website design and everything.
What happened
Two hours later, the next design prompt asked to “make the design the website a whole lot better”.
Stronger prompts · 25 Oct 2026 · excerpts
Act as a Principal Frontend Engineer, Senior UI/UX Designer, Creative Developer, and Software Architect with extensive experience building award-quality interactive developer portfolios.
Accuracy requirement: These are initial content references, not permission to invent achievements. Confirm details against supplied project files, README documents, CV, and other reliable materials. Never fabricate job titles, employment, numbers, credentials, or project results.
Both modes MUST be powered by the same structured portfolio data.
A recruiter with zero terminal experience can browse the entire portfolio easily.
What happened
The first delivery had both modes on one typed data model, 23 passing unit tests and a passing production build. The accuracy rule meant project claims were checked against real files, and unverifiable ones were flagged instead of invented.
| Criterion | Weaker | Stronger |
|---|---|---|
| Concrete goal | no | yes |
| Context | no | yes |
| Constraints | no | yes |
| Acceptance criteria | no | yes |
| Verification | no | yes |
| Scope | no | yes |
Round 02
Asking for a better design
Push the visual design further.
Weaker prompts · 25 Aug 2026
Please make the design the website a whole lot better and push it to github, you can do something crazy 3d animations or anything related to computer engineering, cloud engineering, software engineering etc. Make users be amazed by your capabilities.
What happened
Followed by prompts to change the colour theme and add 3D image effects, then three more prompts to tune the 3D globe (spin speed, drag, sensitivity, extra outlines).
Stronger prompts · 26 Oct 2026 · excerpts
Remove the top navigation completely.
On first opening Terminal Mode, users should see only a minimal terminal interface and the active command prompt.
Specifically avoid: * Identical rounded cards arranged in grids. * Glowing borders around every component. * Generic Lucide icons beside every heading.
Do not prioritize adding more UI. Prioritize making the existing UI exceptional.
What happened
Delivered in one pass, then checked against the brief's own validation list: no navbar, a fullscreen terminal opening to a bare prompt, no project icons, and the approved sections left intact.
| Criterion | Weaker | Stronger |
|---|---|---|
| Concrete goal | no | yes |
| Context | no | yes |
| Constraints | no | yes |
| Acceptance criteria | no | yes |
| Verification | no | yes |
| Scope | no | yes |
Round 03
“Don't make it look AI-generated”
Make the site look deliberately designed, not generated.
Weaker prompts · 5 Oct 2026
Next, I want you to change the whole font of the website to make it "Inter". I do not want the website to be looking very AI generated. (next prompt) yes, tone down the parts that look AI-generated
What happened
Deciding what “looks AI-generated” was left to the model, so there was no rule to check the result against.
Stronger prompts · 26 Oct 2026 · excerpts
The entire website must use only the following three solid colors:
These are the only permitted UI color values.
- NO gradients. - NO colored glows. - NO neon cyan. - NO purple.
Use font weight, spacing, scale, borders, decoration, and contrast to create hierarchy instead of introducing additional colors.
What happened
The taste became a rule, and the rule became an automated test: the build fails if any color other than the three, a gradient, or a palette class appears in the source.
| Criterion | Weaker | Stronger |
|---|---|---|
| Concrete goal | yes | yes |
| Context | no | yes |
| Constraints | no | yes |
| Acceptance criteria | no | yes |
| Verification | no | yes |
| Scope | no | yes |
Round 04
Reporting a bug
Get a layout or behaviour bug fixed.
Weaker prompts · 27 Aug 2026
Look at the attached picture. The picture shows that they are all right beside each other with the red boxes. Please fix that. Make it also that it works with Mobile devices.
What happened
The fix depended on reading a screenshot. The prompt described neither the expected layout nor how to check it.
Stronger prompts · 26 Oct 2026 · excerpts
The prompt correctly updates to `~/projects`, but `ls` still displays the root-level directory contents.
After entering `projects/`, executing `ls` must show the available projects instead of the root directory.
Fix the actual source of the problem rather than patching the screenshot's specific example.
Create or update automated tests covering:
| `cd projects && ls` | Project directories |
What happened
The root cause was found (an empty path resolved to home instead of the working directory), fixed once for every command, and covered by 13 navigation tests taken straight from the brief's table.
| Criterion | Weaker | Stronger |
|---|---|---|
| Concrete goal | yes | yes |
| Context | no | yes |
| Constraints | no | yes |
| Acceptance criteria | no | yes |
| Verification | no | yes |
| Scope | no | yes |
A short prompt that worked
Weaker prompts · 30 Sep 2026
There is a bug, when I move my mouse from top to bottom, my avatar looks down to up which is the opposite. It is supposed to be the same following the mouse pointer. But when I move from left to right, my avatar looks at it correctly.
Short, but it states the actual behaviour, the expected behaviour, and what already works. It was resolved within two more prompts (“perfect thank you, that fixes it”). Length isn't what made the briefs work; specificity is.
What I took away
- 01Define done“Outstanding” and “amazing” can't be checked. A definition of done and a validation list can, and the model will check them for you.
- 02Turn taste into rules“Don't look AI-generated” became three colors and a list of banned patterns, a rule precise enough to enforce with a test.
- 03Give context and limits up frontAudience, source material, and a no-fabrication rule shaped the first delivery instead of being discovered over weeks of follow-ups.
- 04Report bugs as actual vs expectedCurrent behaviour, expected behaviour, and the cases to test. That works in two sentences or two pages.
Limits of this comparison
This compares outcomes, not a controlled experiment. The two sites were built at different times, with different stacks and different scope: the earlier site added a 3D globe, a voxel avatar, a live chess replay and a custom domain, which account for some of its prompts. The scores are the author's own assessment against the six criteria.