How to use AI to write project case studies without violating NDAs or inventing results

The Case Study Conundrum

Every freelancer, agency founder, and software engineer knows the dread of portfolio building. You have done the brilliant work, but now you have to package it into a sleek, narrative-driven case study. Naturally, you turn to an LLM. It is fast, it structures arguments beautifully, and it can turn messy project notes into polished paragraphs in seconds.

But this is where the ethical wheels tend to fall off. If you aren’t careful, a poorly structured prompt will have the LLM hallucinating hockey-stick growth charts in a tick, or worse, leaking your client’s proprietary trade secrets to public training pools.

Using platforms like Claude or OpenAI to draft portfolio content is not inherently lazy or dishonest. However, there is a massive difference between using an AI to organise your thoughts and letting it invent a fictional version of your career. Here is how to navigate the blurry boundary between efficient copywriting and outright fabrication.

1. The NDA Trap: Sanitising Your Data Before You Prompt

The first ethical failure of AI-assisted case studies happens before you even hit enter. If you paste raw, unredacted Slack conversations, database schemas, or internal client presentations into a commercial chatbot, you are highly likely violating your Non-Disclosure Agreement (NDA).

Even if you use enterprise-grade models with zero-retention policies, pasting sensitive data into external servers remains a compliance hazard. You need a strict sanitisation protocol.

• Abstract the brand: Turn "HSBC" into "a major multinational banking institution."
• Obfuscate the tech stack details: Instead of detailing the exact proprietary API architecture, describe it conceptually ("a legacy microservices architecture").
• Mask the metrics: If you improved load times by 42%, you can state that. But if you are detailing raw revenue numbers ($4.2M in Q3), convert them to percentages or relative scale metrics before prompting.

If you are using Anthropic’s models and want to ensure your data isn’t being used for training, consult the Claude Support Hub to review their latest enterprise privacy toggles.

2. The Metric Myth: Preventing LLM "Extrapolation"

LLMs are trained to please. They want to give you a satisfying story arc. In a classic case study narrative, that means the hero (you) must achieve spectacular, world-changing results. If your input notes say "we launched the beta and got some decent initial user feedback," a helpful LLM might rewrite that as: "resulting in unprecedented user adoption and an immediate lift in engagement metrics."

This is not just polishing; it is lying.

To prevent this, you must explicitly constrain the model in your system prompt. Use our prompt generator to build structured system instructions, but always include a variation of this ethical guardrail:

> "Do not extrapolate, assume, or invent any quantitative metrics or qualitative outcomes. If a specific result is not provided in my source notes, state only the direct output of the work (e.g., 'the system was deployed to production') without adding adjectives like 'highly successful' or 'transformational.'"

If you did not measure the exact conversion lift of that UI redesign, do not let the AI guess it. Talk about the structural improvements, the reduction in user friction points, and the cleaner code quality instead. Authentic, moderate metrics are far more believable to a seasoned hiring manager than a series of suspiciously clean "300% growth" claims.

3. Dissecting Your Actual Contribution

In team projects, case studies can easily gloss over who actually did the heavy lifting. When you ask an LLM to "write a case study about our team’s migration to Postgres," the model will naturally write the story from a first-person perspective, making you look like the sole architect of the entire database cluster.

This is a rapid path to getting caught out in a technical interview. To keep things ethical:

• Define your role explicitly: Tell the LLM: "I was one of three frontend developers. My specific responsibility was rewriting the state management system using Redux Toolkit. Do not credit me with the backend API design or the devops pipeline."
• Credit the team: Acknowledge collaborative efforts. It actually makes you look like a better colleague. An ethical case study reads: "While the platform team handled the Kubernetes deployment, I focused on..."

4. The "Human-in-the-Loop" Verification Checklist

Never publish an AI-generated case study without a thorough human pass. Treat the LLM’s output as a rough first draft that needs aggressive fact-checking. Before you publish, run your draft through this quick checklist:

1. The NDA Test: If the client read this, would they be surprised or angry about any of the details revealed?
2. The Verification Test: If a future employer asked for proof of the metrics listed in this case study, could I provide it without sweating?
3. The Vocabulary Test: Does this sound like a real human explaining their work, or does it use generic AI adjectives like seamless, revolutionise, paradigm shift, and testament? (If it does, strip them out immediately. You can find practical definitions of these overused terms in our glossary).

By treating the LLM as a structural editor rather than a creative fiction writer, you can build a stunning, professional portfolio that highlights your genuine skills—without sacrificing your integrity.