The prompt engineering toolstack has matured significantly. This is an honest assessment of what's worth learning — not a sponsor list — organised by how you'll actually use these tools on the job.
LLM APIs: The Foundation
Every prompt engineer needs fluency with at least two major LLM providers:
- OpenAI API — GPT-4o and o-series models. The most widely deployed in UK product companies. Strong function calling, structured outputs, and the Assistants API. The de facto standard for most commercial use cases.
- Anthropic Claude API — Claude 3.5 Sonnet and Haiku. Favoured for document processing, long-context tasks, and applications where instruction-following precision matters. Growing share in UK enterprise deployments.
- Google Gemini API — Gemini 1.5 Pro and Flash. Strongest for multimodal tasks and very long contexts. Less commonly deployed as a primary model in UK product companies but growing in enterprise Google Workspace integrations.
Know the pricing models, context window limits, and key capability differences for each. Interviewers will ask which you'd choose for a given use case and why.
Prompt Management and Versioning
As prompts become production assets, you need version control, testing, and deployment workflows for them — just like code.
- LangSmith — The most complete prompt management platform. Integrates tightly with LangChain but works standalone. Offers prompt versioning, run tracing, and evaluation tools. Widely adopted at UK product companies using LangChain. Has a free tier.
- PromptLayer — Lighter-weight option. Good for teams not already using LangChain. Provides a clean interface for managing prompt versions and tracking usage across API calls.
- Helicone — Focuses on observability and cost tracking. Sits as a proxy in front of your LLM API calls. Particularly useful when you need visibility into per-feature costs at scale.
Evaluation Frameworks
This is the most important category. Without reliable evaluation, you're guessing.
- Promptfoo — Open-source, free, and well-maintained. Define test cases in YAML, run evaluations from the CLI or CI pipeline, and get structured results. In our view, this is the best starting point for anyone building a prompt evaluation practice. Strongly recommended for portfolio projects.
- DeepEval — Python-based evaluation framework with a growing set of built-in metrics (answer relevancy, faithfulness, hallucination detection). Good for RAG pipeline evaluation. Free tier available.
- RAGAS — Specifically designed for evaluating RAG (Retrieval-Augmented Generation) pipelines. If you're working on document Q&A or knowledge base products, RAGAS provides the most relevant metrics.
Evaluation is the highest-value skill
Any engineer can write a prompt. The value in prompt engineering comes from measuring quality reliably and iterating with data. Invest your learning time here before anywhere else in the toolstack.
Orchestration
- LangChain — The most widely-known orchestration framework. Abstractions for chains, agents, and tool use. Worth knowing because it appears in many job descriptions. Main criticism: heavy abstractions that can obscure what's actually happening. Learn it, but also know how to build equivalent pipelines in plain Python.
- DSPy — Takes a different approach: instead of hand-writing prompts, you define the pipeline programmatically and let DSPy optimise the prompts automatically against an evaluation metric. Excellent when you have good eval data and want to find optimal prompts systematically. Growing adoption at research-oriented companies.
Playground Tools for Experimentation
- OpenAI Playground — Fast iteration on system prompts with full control over model parameters. The standard starting point for new prompt experiments.
- Anthropic Console — Anthropic's equivalent, with a good prompt generation assistant. Particularly useful for testing Claude-specific behaviour.
- Google AI Studio — Free access to Gemini models with a generous quota. Good for multimodal experiments and long-context testing.
The Minimum Viable Stack for a Job Seeker
You don't need everything above. Here's what to focus on first:
- OpenAI API + Python — the baseline. Get fluent calling the API directly before using any framework.
- Promptfoo — set up your first evaluation suite. This single tool demonstrates more value than any other.
- Git — version-control your prompts and evaluation configs properly from day one.
- LangChain (conceptual) — understand what it does and when you'd use it, even if you don't build your portfolio projects with it.
Once you have the basics, add LangSmith for prompt management and DeepEval for richer evaluation metrics as your projects grow in complexity.
See the full Prompt Engineer career guide
Salary data, skills breakdown, UK companies hiring, and what the career path looks like.
Frequently Asked Questions
What tools do companies actually use?
Most UK product companies use direct LLM API access, a prompt management layer (LangSmith is widely adopted), an evaluation framework (Promptfoo or custom), and LangChain or lighter orchestration for complex pipelines.
Is LangChain still worth learning in 2026?
Yes, with caveats. It appears in many job descriptions and is worth knowing conceptually. But also learn how to build pipelines without it — over-reliance on abstraction layers is a common interview red flag.
How do I choose between tools?
Start with what the team already uses. Building from scratch? Promptfoo for evaluation, LangSmith if you're using LangChain, direct API access plus Python for everything else. Don't over-engineer early.
Are there free options?
Yes. Promptfoo is open-source and free. DeepEval has a free tier. LangChain and DSPy are open-source. OpenAI Playground, Anthropic Console, and Google AI Studio all have free tiers for experimentation.
What tools should I know for a prompt engineering interview?
Be prepared to discuss: the major LLM APIs and their differences; at least one evaluation framework; LangChain at a conceptual level; and how you would version-control and measure prompts. Demonstrated competence with core concepts matters more than deep expertise in every tool.