What replaced traditional prompt crafting
The specific 2023 craft of finding the perfect phrase to unlock extra performance is mostly dead because frontier models have been trained on those very techniques. RLHF, constitutional AI, and reward models have ingested countless Medium posts, LessWrong essays, and Reddit threads about prompt tricks. The tricks are now in the training distribution—expected and no longer effective at moving the needle.
Examples of outdated techniques:
Native JSON mode, strict function calling, and tool use with input schemas replace free-form parsing. The schema becomes the contract; the prompt becomes a detail.
Replaced agent prompt engineering loops. You describe tools with JSON schemas; the model returns structured tool calls; your code dispatches and appends results. No more parsing or scratchpad regex.
What goes into the context window (position, order, compression) matters more than instruction phrasing. Key factors: position bias, retrieval order, system prompt stability, and compression strategies.
If you can't measure the change, you're guessing. Modern loop: create dataset, define graders, run eval suite on every change, ship only when green. Tools: LangSmith, Braintrust, Langfuse, Promptfoo, Inspect.
You need the model to notice when it's wrong and fix it. Pattern: generate output, validate against schema/test/linter, feed error back and retry. Powers Claude Code, Cursor, and Aider.
Creates Pydantic, Zod, or JSON Schema definitions that pin down what "success" looks like before generation begins.
Specifies what tools the model can call, their parameters, return values, and error messages that teach the model.
Manages retrieval, ranking, compression, cache boundaries, and turn eviction in the context pipeline.
Creates datasets with expected properties, defines graders, and calibrates LLM-as-judge against human ratings.
Reads traces (OpenTelemetry GenAI spans, Langfuse trees, LangSmith sessions) to understand and fix agent behavior.
This is essentially "software engineer who works on an LLM feature." The artifact is the system around the prompt, and the prompt shrinks as you tighten schemas, add tools, or move decisions into code.