Designers, PMs, and engineers share one portfolio problem: the case study reads like a STAR-format dump and never surfaces the actual call you made. A hiring manager scans your opening case study in 0-3 seconds and skims for outcomes in the next ten, so a project that buries its one real decision under a process diary loses before it is read. AI can fix the structure if you feed it the bones — the spine of a portfolio story is the decision under uncertainty, not the checklist of activities.
TL;DR
- Lead with the decision, not the process. The spine of every case study is one call you made under uncertainty — name the alternative you rejected and why.
- Write the first 50 words yourself; your voice is set there. Paste the raw decision into AI to structure and tighten, not to invent.
- Use the copy-ready prompt below. It forces a sub-80-word hook, a 4-5 fact context, an explicit rejected alternative, a number, and one honest “what I’d do differently.”
- Claude Sonnet 4.6 produces fewer AI tells than most models on this kind of writing, which matters when recruiters and applicant systems both flag generic copy as of June 2026.
- 3-5 long-form case studies is the consensus floor and ceiling. Past 5 dilutes; under 3 reads thin.
The task
You are writing or rewriting a portfolio project narrative for your site, a deck, or a take-home review. You want a 600-900 word story that lands the call you made, what you almost did instead, and how you would know if you were wrong.
When this is the right job for AI
- The project shipped and has a measurable or qualitative outcome (real users, real numbers, real adoption).
- You can name the one decision that mattered most — not the five micro-decisions that did not.
- You will write the first 50 words yourself before pasting anything to AI. The voice is set in the first 50 words.
- You have a target audience (hiring manager at a specific company stage / function).
If the project did not ship or has no real users, skip it. A portfolio of unshipped work signals a judgment problem about what to include, not a lack of opportunity.
Which model to use
Any current frontier model can do this, but the failure mode is generic, detectable prose, and that failure is expensive on a job application. As of June 2026:
| Model | Plan | Why for this task |
|---|---|---|
| Claude Sonnet 4.6 | Claude Free (limited) / Pro $20/mo | Fewest AI tells in writing-quality comparisons; strong at cutting filler when told to. Workhorse pick. |
| GPT-5.5 | ChatGPT Free / Plus $20/mo | Fast, good at restructuring; nudge it harder against superlatives and triple-adjective phrasing. |
| Gemini 3.1 Pro | Google AI Pro $19.99/mo | 1M-token context if you are pasting a long raw project log to compress. |
The model is a structuring tool, not a ghostwriter. Treat the output as a framework you edit, never as a final draft you paste. See How to use AI to rewrite resume bullets for the same edit-don’t-generate discipline.
What to feed the AI
- The project in 8-10 lines (problem, your role, what shipped, the outcome)
- The one decision under uncertainty — what you almost did instead and why you did not
- The constraint you held (time, budget, headcount, tech debt, brand)
- The cost-of-being-wrong: what would have broken if your call was the wrong call
- The audience and the format (web case study, 6-slide deck, 1-page write-up)
Copy-ready prompt
You are writing a portfolio project narrative for a {senior product designer} portfolio.
Project (8-10 lines):
{paste — problem, role, shipped, outcome}
The decision under uncertainty:
{e.g. "We chose to ship a partial onboarding flow at week 4 instead of a complete one at week 8. We almost waited because data-quality was at risk; we shipped because customer-success was bleeding tickets."}
The constraint I held:
{e.g. "Engineering had 2 weeks committed; brand wanted launch on Monday; no extra QA."}
The cost of being wrong:
{e.g. "If activation didn't move 10%+, we'd have shipped a worse experience for the same load and burned an internal champion."}
Target audience: {hiring manager at a Series B SaaS}.
Format: web case study, 700-850 words.
Write in this structure:
1. One-paragraph hook (under 80 words) — the call I made + the stakes, no jargon.
2. Context — 1 paragraph, 4-5 facts only, no industry primer.
3. The decision — explicitly name the alternative I rejected and why.
4. What I shipped — specific (number of screens, what was cut, what was deferred).
5. The result — number + a second-order outcome (a behavior change, a downstream metric).
6. What I would do differently — one honest sentence, not three.
Voice rules:
- First person, contractions allowed.
- No "I led" if I was the only one on the project. Say "I designed", "I made the call".
- No "stakeholder alignment" unless you can name the stakeholder and the specific friction.
- No "leverage", "utilize", "seamless", or three-adjective phrases.Sample output structure
Hook: A specific call made in one paragraph — “We shipped a partial onboarding flow at week 4 instead of the complete version at week 8.” The stakes in one clause — “customer success was losing 12 tickets a day to drop-off.”
Context: Four facts only — team size, product stage, the constraint, the success criterion. No SaaS primer. Hiring managers do not need to be told what SaaS is.
The decision: Name the alternative (“hold for the data-quality fix”) and why it lost (“the ticket bleed was a known cost; the data risk was an estimated one”). This is the spine of the story.
What I shipped: Concrete artifacts — 3 onboarding screens replaced with a single in-product checklist, a Notion doc renamed as the runbook for CS, the unglamorous meeting where engineering and CS aligned on what counted as “done.”
The result: 28% lift in activation, 9 fewer tickets per day in week 1, plus a behavior change — CS started running their own SQL queries against the activation table.
What I would do differently: One honest sentence — “I would have shown the partial flow to two CS reps before shipping; I assumed they would be on board and that cost me a week of trust-rebuilding.”
How to refine
- Reads as STAR template: ask AI to remove the visible STAR seams. The hook should not say “Situation:” anywhere.
- The decision is vague: AI cannot infer the alternative if you did not name it. Rewrite the input, not the output.
- Too many activities, not enough call: ask AI to count the activities listed and cut anything past three. Activities dilute the spine.
- Sounds boastful: add “no superlatives.
dramatically improvedbecomes the actual number;worked closely withbecomes the name of the function and the specific friction.” - All projects sound the same: rotate the spine — one project is a speed call, one a tradeoff call, one a no-call (where waiting was the move).
Common mistakes
- Hiding the call. The reader cannot tell what you actually decided versus what was given to you.
- Inflating your role on team projects. “I led” when you co-led, “I designed” when you reviewed. Recruiters check references and the inflation surfaces.
- Showing process artifacts (sketches, journey maps) without naming what changed because of them. Process without a decision is decoration.
- No constraint mentioned. A project with no constraint reads as a project that did not happen — every real project has them.
- One narrative for every audience. Tune the hook for the role: PM hiring → constraint and tradeoff; design hiring → craft and rationale; eng hiring → system and cost.
- Pasting AI output unedited. Generic phrasing is the single most common tell, and a 0-3 second scan catches it first.
FAQ
- How many projects should be in my portfolio? 3-5 long form, plus 2-3 thumbnails for breadth. The 3-5 case-study range is the broad 2026 hiring consensus. Past 5 dilutes; under 3 reads thin.
- Which AI should I use to write the narrative? Claude Sonnet 4.6 (Claude Pro, $20/mo) produces the fewest AI tells in writing comparisons as of June 2026; GPT-5.5 and Gemini 3.1 Pro both work if you push harder against superlatives. The free tiers are enough for a single case study.
- Will recruiters know I used AI? They will if you paste it raw. The fix is the same as for any writing: set your voice in the first 50 words, feed the model real decisions, and edit the output line by line.
- Can I use AI to generate the visuals? For diagrams and flow rebuilds, yes. For final hero shots, no — generated visuals are increasingly easy to spot.
- Should every project have a number? Yes for at least 60% of them. Qualitative-only is fine for craft projects; mixed portfolios should not be all-qualitative.
- What about NDA work? Strip identifiers, change names, keep the decision. The structure of the call is rarely confidential.
- Should I include failed projects? Yes — one. A portfolio with zero failures reads as either junior or sanitized.