Oslo Vibe Coding

Session 06 / AI-native engineering

Context
Engineering.

Designing what an AI can see, use, remember, and verify.

Oslo Vibe Coding

The prompt is no longer the product.

PromptWhat you ask
Context systemWhat the model can know and do next
Oslo Vibe Coding

Teaching analogy

Context engineering is memory management.

The job is to assemble the smallest reliable working set for the next decision.

Processor
LLMReasons over what is currently visible
Working memory
ContextInstructions, examples, evidence, state
Storage
SourcesFiles, databases, repositories, web
Input / outputTools connect the model to the world.
Oslo Vibe Coding

One model. A growing information system.

2017Attention

Models relate information across a sequence.

2020Examples

Behavior changes inside the prompt.

2020Retrieval

External evidence enters the answer.

2022Agents

Reasoning connects to actions and tools.

NowContext systems

Selection, memory, tools, and feedback work together.

Oslo Vibe Coding

More context is not more intelligence.

Teaching heuristic

Quality ~ relevance x accuracy x timing x structure

A long window still needs an editor.
meeting_notes_2024.md
old_architecture_options.md
EU customer data stays in the EU.
unused_api_examples.md
deprecated_roadmap.md
Signal
Oslo Vibe Coding

Shape context with four verbs.

WriteSave decisions and state outside the chat.
SelectRetrieve only what changes the next action.
CompressKeep conclusions, evidence, and open questions.
IsolateMove noisy exploration into a separate context.
Optimize the next decision
Oslo Vibe Coding

Find the requirement

It can fit and still be missed.

Important context competes with position, repetition, and believable distractions.

project-context.md / 18,420 tokens
EU customer data must remain in EU regions.
Oslo Vibe Coding

Turn scattered sources into compounding context.

Raw sources

brief.md
interview.md
policy.md
AGENTS.md / schema and maintenance rules

Maintained wiki

privacy.md
auditability.md
decision.md
ingestquerylint

Better questions

Answer with links, conflicts, and reusable decisions.

The synthesis survives the chat.

Oslo Vibe Coding

Keep the lesson. Not the whole transcript.

01Try
02Observe
03Reflect
04Patch
05Retry
Preserve sources / record confidence / review every patch
Oslo Vibe Coding

Give the agent a map, not a manual.

AGENTS.md
specs/What must happen
docs/How the system works
decisions/Why choices exist
tests/How behavior is proved
Focused Next.js 16 evaluation
53
No docs
53
Skill default
79
Skill requested
100
Docs index

Pass rate reported by Vercel. Narrow task set; not a universal result.

Oslo Vibe Coding

The goal is not more context. It is better decisions.

01

Issue only

Fast, but assumptions fill the gaps.

02

Raw dump

More tokens. More competition for attention.

03

Curated map

Relevant files, constraints, and proof.

04

Progressive context

Retrieve, checkpoint, verify, and continue.

Oslo Vibe Coding

Build context that survives the chat.

07:00
01

Fork

Open the Gist with a file-capable agent.

02

Ingest

Add the three immutable Pulse sources.

03

Question

Should check-ins appear anonymous?

04

Inspect

Review links, conflicts, index, and log.

Ask the agent: Recommend an implementation. Cite every source. Expose unresolved conflicts. Do not modify raw files.
QR code linking to Andrej Karpathy's LLM Wiki Gist
Fork the LLM Wiki Gistgist.github.com/karpathy/442a6bf555914893e9891c11519de94f

Keep the fork. Replace Pulse with your own project.