Most prompt engineering portfolios are terrible. They consist of ChatGPT screenshots, vague claims about "improving response quality", and projects with no evaluation data. Here's what a good one looks like — and three specific project ideas with full instructions.
Why Most Portfolios Fail
The core problem: most prompt engineering portfolios demonstrate output, not process. A screenshot of a clever ChatGPT response tells a hiring manager nothing about whether you can build evaluation infrastructure, think systematically about failure modes, or make data-driven decisions about prompt changes. These are the skills that matter in the job. Your portfolio needs to demonstrate them explicitly.
What a Good Portfolio Demonstrates
Systematic thinking with before/after analysis
Show a prompt iteration process: what you started with, what problem you were solving, what you changed, and — critically — evaluation data showing whether the change improved things across a representative test set. Not subjective impressions. Actual metrics.
Evaluation infrastructure, not just outputs
A test suite with 100+ cases, automated scoring, and a repeatable run process is more impressive than any single clever prompt. The most valuable portfolio evidence is that you can build measurement infrastructure — the part that makes changes safe.
A real deployed tool built with an LLM API
Something accessible and functional. It doesn't need to be complex — a focused single-purpose tool (a document summariser, a code reviewer, a structured data extractor) demonstrates end-to-end capability, not just notebook experimentation.
Evidence of understanding failure modes
Deliberately try to break your prompts. Document what you found. Show you understand jailbreak patterns, prompt injection risks, hallucination tendencies, and how you designed your system to handle them. This is genuinely rare in junior portfolios and highly valued by hiring managers.
What hiring managers look for in the first five minutes
Can this person work on something in production? That means: do they test their work, do they understand failure modes, and can they document what they're doing? If your portfolio answers yes to all three, you're ahead of most candidates.
Three Specific Project Ideas
Project 1: Systematic Prompt Evaluation for a Customer Support Use Case
Create a synthetic dataset of 150–200 representative customer queries — a mix of simple questions, edge cases, complaint escalations, and off-topic requests. Write an initial system prompt, run it against the full dataset, and score outputs on multiple dimensions: answer correctness, tone, refusal appropriateness, and response length.
The key deliverable: a documented iteration. Identify the three biggest weaknesses from your eval results, make targeted prompt changes, re-run the evaluation, and show the before/after delta. Use Promptfoo or DeepEval as your evaluation framework. Publish on GitHub with the evaluation scripts, prompt files, and a write-up explaining your decisions.
Project 2: A Production-Ready Structured Data Extraction Tool
Build a tool that extracts structured data from unstructured text — job descriptions, receipts, academic papers, or a domain you know well. The system should accept free-text input, extract it into a defined schema using function calling or structured outputs, validate the output, handle cases where the input doesn't contain expected data, and offer a simple web interface.
Deploy it free on Vercel or Hugging Face Spaces and open-source the code on GitHub. Document your schema design decisions and how you handle ambiguous or incomplete inputs. Structured output extraction is one of the most common real-world LLM tasks — building it reliably demonstrates practical, production-ready skills.
Project 3: A Safety and Failure Mode Analysis
Choose a real LLM-powered product or open-source tool. Systematically probe its prompt design: attempt common jailbreak techniques, try prompt injections, test edge cases that weren't clearly intended. Document every finding, classify failures by type and severity, and propose specific prompt changes to address the most serious issues.
Deliver a structured report: executive summary, findings table with severity ratings, root cause analysis for each issue, and recommended mitigations. This is the kind of thinking that prompt engineers at safety-conscious UK companies — fintech, legaltech, healthcare AI — do every day.
Important: only test tools you have permission to test, or your own prompts. Don't publish confidential system prompts or claim to have found vulnerabilities in commercial products without responsible disclosure.
Where to Publish
The minimum viable portfolio: a well-organised GitHub profile plus a personal blog or Substack with write-ups explaining your thinking. For each project: a clear README, clean code, and a link to the write-up. If you have a live demo, link it prominently in the README header.
LinkedIn is where you'll share with hiring managers. Post write-up summaries with links. AI hiring managers scroll LinkedIn and will find good work when it's clearly written and easy to navigate.
Explore the Prompt Engineer career guide
Full salary tables, skills breakdown, UK companies hiring, and career path guidance.
Frequently Asked Questions
Do I need a GitHub profile?
Yes. GitHub is the standard way to share evaluation scripts and prompt files with hiring managers.
Can I use ChatGPT conversations as portfolio evidence?
Only as illustration within a broader write-up — not as primary evidence of your skills.
How many projects do I need?
Two to three strong projects outperform five weak ones every time.
Should I include failed experiments?
Yes, selectively. A well-documented failure that shows rigorous reasoning is genuinely impressive to hiring managers who know the domain.
What format should my portfolio take?
GitHub repo plus blog write-ups. A live demo adds strong credibility. Optimise for clarity — hiring managers spend 5–10 minutes reviewing.