Wardley Maps connect user needs to the dependencies that deliver them, then show how those components move from custom work toward commodity services.
2
Prompt engineering includes guardrails, chains, agents, private-data access, embeddings, vector databases, models, and the infrastructure underneath them.
3
OpenAI could turn parts of its infrastructure into APIs and compete directly with services that teams are building around its platform.
Summary
Mark Craddock introduces Wardley Mapping as a way to plan strategy around user needs, value chains, dependencies, and the movement of technology from Genesis through custom products toward commodity services. He applies the method to prompt engineering. The resulting map includes prompting techniques, guardrails, chains, agents, private data, privacy-enhancing technologies, embeddings, text splitting, vector databases, search algorithms, cloud infrastructure, and large language models. Craddock uses the map to discuss build-or-buy decisions, investment, technical debt, change, cost per transaction, and the risks of relying on platform providers. He argues that prompt engineering is larger than writing prompts because it includes the surrounding systems and operational concerns. He also warns that OpenAI could expose more of its internal capabilities through APIs and compete with products built on its platform.
Wardley Maps start with user needs and trace the dependencies behind them
Craddock describes Wardley Mapping as a strategic planning tool created by Simon Wardley. A map begins with a user need and then links the components required to meet it. The map runs from Genesis on the left toward Commodities on the right. Computing illustrates the movement: it began as research and mechanical systems, then became a utility. The links show dependencies and the flow of capital, which can include money, trust, or other resources. Craddock uses this structure to make decisions about investment, focus, competition, and whether to build or buy a component.
A Wardley Map can give an LLM the structure behind a technical system
Craddock says the map can be represented as code and supplied to ChatGPT. The model can then understand the components and their links, help build the map, challenge it, or create additional maps. In his example, the chain includes large language models, cloud services, chunking, and text splitters. Encoding the relationships gives the model context about how the system works. Craddock presents this as a practical use of mapping, since the map can support strategy discussions as well as technical analysis.
The prompt engineering map includes system controls and operational concerns
The simplified prompt engineering map begins with use cases such as answering questions over private data. That requires conversation history and prompting techniques, including what Craddock calls prompting, where the value chain is placed into the prompt to provide context. Below those techniques are guardrails, chains, and agents. Guardrails should enforce the ethical conditions chosen for the service. Agents can query multiple models, tools, external data, websites, and APIs, then return the results to a language model. Private data may require custom agents and metadata about the data.
Private-data systems depend on embeddings, splitting, search, and vector storage
Craddock places embeddings, text splitting, vector databases, search algorithms, and cloud infrastructure underneath applications that use private data. Encrypted embeddings can protect the private layer. He describes fully homomorphic encryption and multi-party computation as ways to analyse data while it remains encrypted. Text must be split according to its format, such as PDFs, plain text, or YouTube scripts. A vector database then stores the vectors, while nearest-neighbour search retrieves related material. He names Supabase, Chroma, Pinecone, and Facebook's similarity-search technology as examples in this part of the stack.
Privacy and cost belong on the same map as model features
Craddock says teams should consider privacy-enhancing technologies, DataOps, MLOps, and FinOps. FinOps matters because large language model costs depend on tokens and the pricing models of providers. He notes that model prices can change, using OpenAI's price reduction as an example. Privacy concerns also affect the architecture. Concrete ML provides encrypted machine learning models that can be queried while encrypted, and Cape Privacy works on protecting and encrypting chat data. The map therefore connects privacy controls and operating costs to the model and application layers.
Some parts of the stack are already moving toward commodity status
The map uses red lines to show expected movement through the evolution axis. Craddock says agents are in the product space and are likely to move toward the commodity space over time. Prompt templates may already be commodities because thousands are available on the internet. Open source models can accelerate movement toward commodity services. He gives examples including OpenAI, Dolly, Alpaca, Llama, Orca, Falcon, and Meta's multilingual speech model. The map is intended to be updated as these technologies and their positions change.
Prompt engineering covers much more than writing a prompt
Craddock rejects the idea that prompt engineering is only a narrow writing task. He calls it prompt crafting when the work is simply writing prompts. The wider map includes computer science work around nearest-neighbour search, data science work around large language models, and MLOps work around infrastructure. The prompt engineer sits above these layers and connects them to the user need. This framing places model choice, retrieval, agents, infrastructure, privacy, and operations inside the same engineering problem.
Platform providers can become direct competitors to products built on them
Craddock warns that OpenAI has a platform and could expose parts of its infrastructure through APIs. If it does, the provider could enter areas where other teams are building services. His point is a strategic risk revealed by the map: a team may build around a capability that the platform owner can later offer directly. The map helps make that dependency visible before deciding where to invest. It also gives teams a way to compare their position with the evolution of the underlying components.
"If they decide at any point to make any of this available for other APIs, they're going to eat your services you're currently thinking about building."09:19
Who should watch
You are deciding whether to build or buy parts of an LLM application and need a way to compare the dependencies.
Your team is building over private data and needs to think about retrieval, privacy, model costs, and operational layers together.
You want to understand how platform providers and changing technology maturity can affect an LLM product roadmap.
