Session 06 / AI-native engineering
Context
Engineering.
Designing what an AI can see, use, remember, and verify.
Session 06 / AI-native engineering
Designing what an AI can see, use, remember, and verify.
No one codes alone.
The shift
Better output starts before the model answers.
A useful mental model
Teaching analogy
The job is to assemble the smallest reliable working set for the next decision.
Load less. Load better. Load when needed.
How we got here
Models relate information across a sequence.
Behavior changes inside the prompt.
External evidence enters the answer.
Reasoning connects to actions and tools.
Selection, memory, tools, and feedback work together.
The model stayed central. The surrounding system grew.
Attention budget
Quality ~ relevance x accuracy x timing x structure
A long window still needs an editor.Capacity tells you what fits. Engineering decides what belongs.
Four moves
Write. Select. Compress. Isolate.
Small experiment
Find the requirement
Important context competes with position, repetition, and believable distractions.
"It fits" is not the same as "it will be used."
LLM Wiki
The synthesis survives the chat.
RAG retrieves. A maintained wiki compounds.
Adaptive context
Useful memory is selected, tested, and revisable.
Vibe coding practice
Pass rate reported by Vercel. Narrow task set; not a universal result.
Route the agent to evidence before it trusts memory.
Context ladder
Fast, but assumptions fill the gaps.
More tokens. More competition for attention.
Relevant files, constraints, and proof.
Retrieve, checkpoint, verify, and continue.
Shape the context for the next action.
Seven-minute workshop
Open the Gist with a file-capable agent.
Add the three immutable Pulse sources.
Should check-ins appear anonymous?
Review links, conflicts, index, and log.
Keep the fork. Replace Pulse with your own project.
No one codes alone. oslovibecoding.tech