Prompts
Prompt engineering didn't die. It got absorbed.
The obituary misses the address
Prompt engineering didn't die. It got absorbed.
“Prompt engineering is dead” usually points to something real. Newer models often need less ritual than earlier chat systems did. Stock phrases, elaborate roles and collections of supposed magic words are less reliable than clear instructions and good evidence. But the useful discipline was never the ritual. It was expressing intent precisely enough that a model could act on it.
That discipline remains. What changed is its address. Intent can now live in the request you type, but also in standing instructions, tool contracts, examples and the checks that judge the result. Retrieved context supplies the evidence those instructions act on. Prompt engineering did not disappear. It was absorbed into the larger system around the model.
So when people say the skill vanished, what they're really reacting to is a change of address. The sentence stopped being the whole lever.
Five places intent can live
The message box became one address among several
In early chat interfaces, most intent had to fit inside the next message. As AI systems gained more durable structure, the instruction could be placed where it would do the most useful work.
A task prompt describes what is different this time. A standing instruction carries rules that should apply across tasks. A tool schema, the contract describing what a tool accepts and returns, makes the requested action explicit. Permissions and server-side validation determine what the tool will actually allow. Examples show what good work looks like when prose rules are ambiguous. An evaluator checks whether the output actually met the requirement. The same intent may appear in more than one place when the consequence justifies the reinforcement.
The design question is no longer only “How should I phrase this?” It is also “Where should this rule live, who maintains it, and how will we know it worked?” That is a broader engineering problem, but it still depends on the original skill of saying what you mean.
Intent moved from one message to a system
The immediate prompt is now only one address where behaviour is specified.
Visible task message
- Current outcome What this run should produce.
- Local constraints What is special about this instance.
- Current question The decision the model must help resolve.
same intent, more addresses
Standing operating contract
- Instructions and examples Durable behaviour and representative cases.
- Tool schemas and permissions What actions mean and which are allowed.
- Evaluators and escalation How failure is detected and authority returns to a human.
Prompt quality still matters because the task message selects and applies the wider system.
A worked example
One status report, five forms of intent
Suppose a team uses AI to prepare a weekly project report. The task prompt says, “Summarise Project North for the week ending Friday.” Standing instructions define the audience, require facts to be separated from inference and forbid personal customer data. The reporting tool requires a project identifier and date range. Permissions restrict which projects are accessible, server-side validation rejects a period outside the requested report, and the evaluator checks that cited dates match the target week.
Two approved examples show the expected shape: outcome, current risk, named owner and next decision. Finally, an evaluator checks that every risk has a source and an owner before the report can be sent. The task prompt is still important because it names this project and this week. The durable instructions, tools, examples and checks prevent the team from rebuilding the same definition of a good report every Friday.
This is the lineage I find useful. Prompt craft taught us to express the request. Context engineering taught us to choose the material placed in front of the model. Harness engineering gives those choices durable places to live and checks how they behave together. These are overlapping layers, not clean historical eras that every team entered on the same date.
Where the task prompt still matters
Durable rules cannot predict every question
Ad hoc prompting still matters when the task is novel, exploratory or unusually sensitive. A standing rule can define how reports are normally written, but it cannot know why this week's board discussion changes the emphasis. A good task prompt supplies that local judgement. It can also ask for alternatives, expose uncertainty or deliberately depart from the default when an authorised person explains why.
The boundary cuts both ways. Moving an instruction into the harness does not make it correct. Durable ambiguity can produce consistent mistakes, and examples can teach accidental patterns. Instructions need owners, tests and revision when the work changes. Some intent belongs in a shared organisational memory; some should expire with the task.
That is the student lesson for me: each correction is a chance to ask whether I learned something local or something reusable. If it is local, improve the next prompt. If it is reusable, give it a durable address and an evaluator. The craft survives in both decisions. Prompt engineering did not die. It became the discipline of placing clear intent where the work can keep learning from it.