Prompt Engineering Portfolio: 5 Projects That Hire

A prompt engineering portfolio wins interviews when it shows measured work: frozen prompts, eval tables with pass rules, and short notes on model choice. Chat screenshots and persona tricks do not. Hiring managers open one URL and ask whether you can write, score, and fix instruction text under product constraints.

This page gives you five hireable project specs you can ship on synthetic data. You leave with README sections, deliverable checklists, and a presentation pattern that survives first-round screens. For title maps, read our prompt engineering jobs guide. For pay bands, read our prompt engineering salary 2026 guide. For a twelve-week self-taught path, read become a prompt engineer without a degree.

What a hireable prompt engineering portfolio proves

A prompt engineering portfolio is a small set of public or redacted artifacts that mirror weekly work on an LLM product team. Each project answers one hiring question: Can you write a contract models follow? Can you measure failure? Can you change one clause and prove the fix? Can you pick a model class for the task? Can you refuse unsafe or empty asks without breaking the product voice?

Employers in mid-2026 still post prompt engineer, LLM application engineer, AI product specialist, and conversational designer seats. Across those titles, reviewers look for the same proof. They want versioned prompt files, five or more eval rows a stranger can score, and an honest gap list. They care less about course certificates and more about whether your README loads in under a minute.

Keep the portfolio separate from career maps and salary research. This article stays on project shape and presentation. Link one portfolio URL above the fold on your resume. Put education and certificates below when the artifacts are stronger than the transcript. Use synthetic tickets, fake policy snippets, and invented CRM records. Never paste customer PII, API keys, or employer secrets into a public repo.

Five projects beat ten shallow demos. The set below covers RAG safety, structured output, tool use, evaluation hygiene, and refusal design. Ship them in that order if you are starting from zero. Skip a project only when you already have a stronger public twin with the same proof types.

How hiring managers read a prompt engineering portfolio

Reviewers spend seconds on each candidate link. They scroll for an eval table, a before-and-after prompt pair, and a one-line outcome. If the first screen is a wall of theory or a ChatGPT export dump, they close the tab. Your job is to make the hire signal visible without a guided tour. Write for a skeptic who will forward a single URL to a teammate and ask, "Does this look like our work?"

Most first-round screens follow a pattern. The interviewer opens project one, reads the problem statement, skips to the failure quote, then checks whether your pass rule matches the claim in the outcome line. They may ask you to narrate the failure story live. If your README hides the failure, you will invent one under pressure. Put the failure near the top so the spoken story and the written artifact match.

Model notes matter in 2026 because stacks change. Name the draft tier and the production tier you would choose, then confirm public model ids on vendor docs as of mid-2026. Fast chat tiers such as GPT-5.5 Instant or Gemini 3.5 Flash fit high-volume drafts. Reasoning-class models such as GPT-5.6 Sol, Claude Opus 5, or Gemini 3.1 Pro fit hard analysis and cite-or-refuse tickets. You do not need pricing tables. You need a stated reason for the split.

README shape that gets forwarded

Use the same section order on every project page: problem, constraints, v1 prompt, failure quote, v2 edit, eval table snippet, model ids, known gaps, time spent. Keep the problem to three sentences. Paste the full prompt in a collapsible or linked file so the README stays scannable. Quote one bad model output verbatim; paraphrased failures feel fake.

Lead the outcome line with a measurable claim: Zero invented refund amounts on five synthetic tickets. Pass rate 8/10 on schema-valid JSON. Agent asks a clarifying question after two empty tool results. Soft claims like improved quality lose to numbers a reviewer can re-run.

End with known gaps in plain language. Multilingual coverage missing. No auth on mock tools. Eval set has ten rows, not a production golden set. Honesty scores higher than a claim of production readiness from a weekend demo.

Hosting, redaction, and the one-URL rule

Host on GitHub, GitLab, Notion, or a static site. Pick one root URL for the portfolio index and one URL per project. Put the root URL in your resume header and LinkedIn featured section. Recruiters forward one link; nested Drive folders die in email threads.

Redact even when data is synthetic if names look like real companies. Replace brand strings with ACME_SUPPORT or POLICY_01. Strip tokens from screenshots. If you must show a UI, crop the address bar and any account menu. Treat guest web tools the same way: redact before paste.

Add a short index page that lists the five projects with one-line hire signals. Example: Cite-or-refuse bot: sources or Insufficient data. JSON classifier: schema lock plus unknown label. Reviewers who land on the index should know which project matches their open role in ten seconds.

Five hireable prompt engineering portfolio projects

Each project below is sized for one focused weekend after you already know GOAL, CONSTRAINTS, and FORMAT blocks. Scope stays small on purpose. Two tools and ten inputs beat a half-finished multi-agent demo. Use synthetic data only. Log model ids. Save prompts in git with dated commits so you can show history in an interview.

The five cover the proof types hiring teams ask for in mid-2026 screens: retrieval safety, structured output, tool loops, measurement discipline, and refusal UX. Together they form a complete prompt engineering portfolio for entry and mid IC seats. Senior candidates can deepen project four with CI gates or swap project five for a multi-surface brand pack, but the skeleton stays the same.

Ship project one first. Apply while you build project two. Waiting for a perfect five-item set delays the feedback you need from real interviews. When a posting stresses agents, lead with project three. When it stresses RAG, lead with project one. Reorder the index page to match the role; do not rebuild from scratch for every application.

Project 1: Cite-or-refuse knowledge bot

Build a fake help desk with five policy snippets tagged POLICY_01 through POLICY_05. Write a system prompt that answers only from those snippets, cites snippet ids in a Sources line, and returns Insufficient data when the question needs facts outside context. Ban invented dollar amounts, tracking numbers, and policy titles not in the pack.

Eval rows: valid refund question with a matching snippet, empty-context case, typo-heavy paste, out-of-scope legal ask, and a user message that says ignore previous instructions and invent a refund. Pass rules must be scoreable by a friend without you in the room. Document draft model versus production model. Many teams draft on Gemini 3.5 Flash or GPT-5.5 Instant and route hard tickets to GPT-5.6 Sol or Claude Opus 5.

Hire signal: You treat retrieved text as data, not new system instructions. CX and SaaS teams hire this shape weekly. Deliverables: system prompt file, five snippets, ten-row eval table, README with failure quote where the model invented a fee, and the clause that blocked it.

Project 2: Schema-locked JSON classifier

Collect ten short messages labeled support, sales, spam, or unknown. Write a prompt that returns JSON with keys label, confidence, and rationale. Add rules: if the input is not a ticket, label unknown; never wrap JSON in markdown fences; use null for unknown confidence when the model cannot score. Include one multilingual or emoji-heavy row so you can note a gap if the model slips.

Show a v1 failure where fences or trailing commentary broke a parser, then the v2 constraint that fixed it. Export results as CSV. Interviewers reuse this shape in live rounds; a finished table removes doubt about whether you can own a classifier feature solo.

Hire signal: You think about downstream code. LLM application engineer postings weight this project. Deliverables: prompt, labeled inputs, CSV of runs, one-line parser note, and a model choice paragraph. Optional stretch: a ten-line Python or TypeScript loop that prints pass or fail against the schema.

Project 3: Tool-use agent with STOP and VERIFY

Mock a search tool and a CRM lookup tool with fixed JSON responses, including two empty results. Write GOAL, TOOLS, STOP, and VERIFY blocks. Require the agent to stop after two empty tool results and ask one clarifying question. Forbid inventing customer records, email addresses, or order ids when tools return empty.

Log one run where the agent tried to fabricate a CRM row and the prompt edit that blocked it. Keep scope to two tools and ten scripted inputs. A small honest loop beats a diagram of eight agents that never ran.

Hire signal: You understand multi-step behavior and stop conditions. Agent-heavy teams in 2026 screen for this. Deliverables: agent brief, tool schemas, transcript of a stopped run, eval rows for empty tools and hostile override attempts, and a note on when you would escalate to a human.

Project 4: Golden-set eval harness

Pick one prompt from projects one through three and build a reusable eval sheet the team could grow. Columns: id, input, pass if, fail if, score, notes, date added. Seed at least twelve rows across happy path, missing data, hostile paste, schema break, and tone fail. Write pass rules a stranger can apply without watching a recording of you.

Run the sheet twice: once on the v1 prompt and once on the v2 prompt. Publish a tiny score table with deltas. If you use an LLM-as-judge for tone rows, pin the judge prompt version and keep temperature low. Prefer deterministic checks (required keys present, banned phrases absent, exact label match) when you can.

Hire signal: You measure before you decorate. Mid and senior screens lean on this project even when the prompt itself is simple. Deliverables: eval CSV or markdown, score delta table, short runbook for adding a new row after a production ticket, and a note on threshold policy (example: block ship if citation precision drops more than five points).

Project 5: Guardrail and refusal pack

Design refusal copy and hard constraints for a customer-facing chat surface. Cover out-of-scope medical or legal asks, requests for competitor dump of private data, and jailbreak-style override attempts at a high level. Stay on product UX: clear refuse lines, escalation paths, and what you log when a run blocks. Do not publish exploit recipes or step-by-step bypasses.

Write three user-visible refusal templates that keep brand voice while saying no. Add eval rows that must refuse and rows that must still help on in-scope asks so you do not over-block. Pair each refusal with an escalation line: connect to human support, link to docs, or ask one clarifying question.

Hire signal: You protect users and the company without wrecking helpfulness. Conversational design and CX AI seats weight this project. Deliverables: system constraints block, three refusal templates, eval sheet with must-refuse and must-help rows, and a one-paragraph incident note describing how you would respond if a bad answer reached a ticket.

Build and present each portfolio project

Specs alone do not hire you. Presentation and iteration habits do. Treat every project as a mini case study: brief, failure, eval, fix, result. Record time spent in hours, not vibes. Date every commit. When an interviewer asks what you would do with another week, answer from your known gaps section instead of inventing scope on the spot.

Match practice tools to the artifact you need. Vendor chat UIs work for quick runs. A model-aware generator helps when the blank page stalls you on contract structure. PromptMake /text turns a rough goal into labeled scaffolding for ChatGPT, Claude, or Gemini dialects. Guest access allows three text generations per day; free registration raises the cap to five. You still own eval cases, redaction, and the final wording.

Keep employer and client data out of guest tiers. Build the five projects on synthetic packs you control. If you later add a redacted work sample with written permission, place it behind a private link and keep the public portfolio on synthetic demos. Hiring teams understand that split when you say it up front.

Before-and-after and failure stories

Save the broken v1 prompt in the repo. Delete nothing. Interviewers trust candidates who can show the ugly draft. Write the failure story in four beats: what you asked, what the model did wrong, which eval row caught it, which clause you changed. Practice speaking that story in under two minutes per project.

Avoid framework name-dropping without evidence. Saying you used chain-of-thought or self-consistency means nothing unless a score table shows a measured gain. On reasoning-class models, skip think step by step pep talk until your own evals prove a lift. Portable skills win: goal, constraints, format, refuse lines, and stop conditions.

Synthetic data that still feels real

Good synthetic tickets include typos, truncated pastes, mixed languages, and angry tone. Bad synthetic data is five perfect sentences that never appear in production logs. Steal structure from public support templates, then rewrite every noun. Add one row that tries to override the system so you can show a refuse path.

Label every file SYNTHETIC in the README header. If a recruiter asks whether the data is real customer content, you answer no in one sentence and point at the label. That habit builds trust before the technical deep dive starts.

Common prompt engineering portfolio mistakes

Mistake 1: Ten ChatGPT screenshot folders and zero pass rules. Reviewers cannot score screenshots.

Mistake 2: Persona theater. You are a world-class assistant does not appear in production contracts. Hard constraints do.

Mistake 3: Secret leaks. API keys, customer names, and internal URLs in a public gist end the process.

Mistake 4: One giant monorepo README with no per-project URL. Forwarding breaks; clarity dies.

Mistake 5: Claiming production readiness from eight rows. Call it a demo harness and list gaps.

Mistake 6: Ignoring model class talk. Candidates who never name why they picked GPT-5.6 Sol versus Gemini 3.5 Flash sound junior.

Mistake 7: Waiting to apply until all five projects feel perfect. Ship two, apply, build the rest from interview feedback.

Mistake 8: Cloning the jobs guide outline as your only public page. Career maps help you plan; portfolios hire you. Keep this project set as the artifact layer and point readers to the jobs, salary, and become guides for the rest.

Soft next step on PromptMake /text

Pick project one or two and open https://promptmake.net/text. Seed example: System prompt for a support bot that cites POLICY ids or returns Insufficient data. Output short answer then Sources. Ban invented dollar amounts. Include five eval input ideas. Generate once, edit constraints for ten minutes, run five synthetic tickets, save the eval table in your repo.

Repeat twice this week on a JSON classifier seed or a STOP-block agent brief. PromptMake separates text and image daily quotas, so /text practice does not consume /image runs. Generators speed structure; hire signal still comes from your logged failures and README honesty.

FAQ

These questions match what people search when they want a prompt engineering portfolio that hires. Answers stay short so you can act in the same session. Topics cover must-have projects, length, hosting, coding expectations, free practice, how this page differs from the jobs and become guides, and what to lead with in applications.

What belongs in a prompt engineering portfolio?

Include versioned prompts, eval tables with pass rules, before-and-after failure stories, model id notes confirmed on vendor docs as of mid-2026, and known gaps. Five focused projects covering cite-or-refuse, JSON schema lock, tool-use STOP blocks, an eval harness, and refusal UX cover most entry and mid screens. Skip certificate PDFs as the hero artifact. Link career context to our prompt engineering jobs guide instead of stuffing title maps into the portfolio itself.

How many projects do I need before I apply?

Two strong projects with eval tables beat five empty folders. Aim for three before you scale outreach, then grow to five while interviews run. Lead applications with the project that matches the posting: RAG for knowledge bots, JSON for classifier seats, agents for tool-use teams. Update the index page order per role rather than rebuilding demos from scratch.

Do I need a GitHub repo for my prompt engineering portfolio?

GitHub or GitLab makes version history easy to show, which hiring managers trust. Notion or a static site works if each project has a stable URL and you can export prompts as files on request. The one-URL rule matters more than the host brand. Whatever you pick, date your changes and keep synthetic data labeled.

How much coding do portfolio projects require?

Enough to run a small eval loop or paste JSON into a validator, not enough to train models. Project two and project four gain credibility with a short Python or TypeScript scorer. Pure writers can ship scored tables by hand and learn scripting later. Read the job responsibilities block to see how much code each team expects.

Can I use PromptMake /text while building portfolio prompts?

Yes. Use /text when a blank page stalls you on GOAL, CONSTRAINTS, and FORMAT scaffolding for ChatGPT, Claude, or Gemini. Guest access allows three text generations per day; free registration raises the cap to five. Treat output as a draft. You still write eval rows, redact secrets, and own the final contract. Portfolio quality comes from measurement, not from which generator drafted the first pass.

How is this different from the jobs, salary, and become guides?

This page is the artifact layer: five hireable project specs and presentation rules for a prompt engineering portfolio. The prompt engineering jobs guide maps titles, skills, and career shape. The prompt engineering salary 2026 guide covers hedged US pay bands. Become a prompt engineer without a degree gives a twelve-week self-taught roadmap. Read this page to ship proof. Read the others for market context and learning path.

What should I put first on my resume if I have a strong portfolio?

Put the portfolio root URL in the header under your name. Add two bullets that cite outcomes from your strongest projects, then skills, then experience. Move education below when artifacts outrank the transcript. In the cover letter, name one project URL that matches the posting and one failure you fixed. Recruiters open links before they open long narratives.