Lineage
Prompt, context, harness: three layers that accumulate.
A working model, not a universal history
My attention moved from the request to the system around it
I use three eras to make sense of how my own AI practice changed: prompt, context and harness. This is not a claim that the whole field crossed three borders on the same dates. The layers overlap, and different teams still meet them in a different order. It is a record of where I found the next useful improvement after the previous one stopped being enough.
At first I concentrated on the request. Then I learned that a clear request fails when the model lacks the right evidence. Later I learned that choosing that evidence by hand for every run does not scale. My attention moved outward from one message to the working environment around it. The earlier skills did not disappear. They became parts of a larger system.
Leverage moves outward from the prompt
As the task becomes repeatable and consequential, more value sits in the operating system around the model.
- Clear prompt Defines the immediate job.
- Selected context Supplies evidence and decisions.
- Tool access Connects reasoning to inspectable action.
- Permissions and stopping Contains reach and consequence.
- Evaluation and learning Rejects failures and improves the next run.
The prompt remains necessary. The repeatable advantage increasingly lives around it.
Layer one: the prompt
First, make the job clear
Suppose the recurring task is a weekly client update. “Write an update” leaves too much undecided. A better prompt names the audience, the purpose and the shape of a good answer: summarise progress, identify decisions the client must make, flag risks and keep the note under four hundred words.
That precision still matters. Modern models are generally better at interpreting ordinary language than earlier ones, but they cannot reliably satisfy a specific standard they were never given. Prompting is the layer where a vague intention becomes an inspectable job. It gives the model a target and gives the person reviewing the output something against which to judge it.
Layer two: the context
Then, put the right evidence in the room
The prompt can be excellent and the update can still be wrong. The model needs this week's project status, the last agreed actions, current measures and any new risk. If it receives an old plan or a speculative note alongside the approved record, fluent writing may hide a weak account of what actually happened.
This is the work of context engineering: selecting the evidence a task needs, distinguishing current sources from stale ones and leaving irrelevant material out. More context is not automatically more truth. The useful question is whether each item helps the model complete this job and whether its authority is clear.
Prompt and context now work together. The prompt says what the update must do. The context gives it facts worth saying. The limitation is repetition: if a person has to find, check and attach the same classes of evidence every Friday, the quality still depends on someone remembering the routine.
Layer three: the harness
Make the good preparation repeatable
A harness is the repeatable working environment around the model. For the weekly update, it might retrieve the latest approved status, carry forward unresolved actions, exclude documents outside the project, apply the standing update format and check that every claim points to a source. It can draft the note, but a named professional still approves what goes to the client.
The harness does not replace careful prompting or context selection. It encodes those decisions and applies them repeatedly, while adding tools, permissions, memory and checks. This is the practical shift described by harness engineering: improve the environment once, then let future runs inherit the improvement. It still needs maintenance when the task, policy or source systems change.
Graduate a useful prompt into a controlled workflow
Add structure only when repetition, consequence or coordination earns it.
- Prompt One person states the task clearly.
- Context packet Relevant evidence becomes repeatable.
- Tool contract Actions, inputs and side effects become explicit.
- Evaluation A check can reject the output.
- Owner and maintenance Someone governs changes and retirement.
Stop adding layers when the next layer costs more than the risk or repetition it removes.
What changed at the centre
Better models raise the ceiling; the layers still accumulate
The model did not hold still while these practices developed. Models became better at interpreting intent, working across longer inputs and using tools. Individual releases can move those abilities unevenly, as the Opus 4.8 evaluation note illustrates. A better model can raise the ceiling and make some old workarounds unnecessary.
But improvement at the centre does not remove the surrounding questions. What are we asking? Which evidence is authoritative? What may the system do? How will we check the result? Prompt, context and harness are therefore more useful as an accumulated stack than as three obsolete eras.
The possible next step is organisational rather than inevitable. Teams may connect shared memory, tools and decision records into a governed intelligence layer. That can spread useful learning beyond one operator, but it also increases the need for ownership, permissions and review. Moving outward increases the leverage and the responsibility together.
The practical diagnostic
Fix the layer where the failure actually lives
When a result disappoints you, diagnose the layer before adding machinery. If the job is vague, clarify the prompt. If the evidence is missing or noisy, repair the context. If the same preparation and checks must be rebuilt every week, make them part of the harness. Then run the task again and see whether the failure changed.
The three layers are not exhaustive. A failure may come from model capability, a defective source system, an unsuitable task or an organisational decision that software cannot repair. The diagnostic should locate the constraint, not force every problem into a preferred framework.
That is the student lesson under this working model: do not defend the era you already understand. Look for the constraint the work is teaching you about now. The professional goal is not a larger AI system. It is reliable work with clear responsibility and less attention lost to rebuilding the same conditions by hand.