Lessons from the Frontier^AI

Curate your context

An average LLM context holds 200,000 tokens.

That means ~250 tokens for each of your 800 endpoints.

If each tool description was about as long as the text on this slide, you've used 1/4 of memory...

just by saying hello.

It's not about less or more.

It's about the right context, carefully chosen.

"We expanded prompts from ~100 words to 2,000+ — but every word was curated.
Explicit column definitions. Precise output formats. Concrete decision rules."

Think in checkpoints, not threads

You don't re-read chapters 1 to 9 before starting chapter 10.

/compact and /clear are your friends.

Summaries and decisions  >  raw conversation history

Training data has a cutoff

Models don't know about your latest library version
or last week's blog post.

The model will confidently use a deprecated API
if it doesn't know the new one exists.

Tell it to web search.

Curate your tools

Tools ≠ utils

We abstract code into small, composable functions.
That's good engineering. It's bad tool design.

"Agents find needles in haystacks — by examining every piece of hay."

Outcomes, not operations

Write agent stories

User stories, but the persona is the agent.

Curate your output

AI output doesn't need to match what the product needs.

Design a translation layer. Include unit tests.

The problem: MUI needs an exact shape

// MUI data grid needs BOTH of these, precisely correct:

columnVisibilityModel: {
  name: true, sku: true, status: true,
  qty: true, internalNotes: false, ...
}

orderedFields: [
  "__check__",        // ← system column (checkbox)
  "name", "sku",      // ← user columns in exact order
  "status", "qty",
  "actions"           // ← system column (buttons)
]

AI was unreliable at producing this — forgot system columns,
wrong ordering, flipped visibility booleans.

The fix: let each do what it's good at

// Semantic — natural for an LLM
{
  identifyingColumns: ["name", "sku"],
  relevantColumns:    ["status", "qty"],
  insightColumns:     ["updatedAt"],
  hideColumns:        ["internalNotes"]
}

// Deterministic — unit tested
mergeAIColumnVisibility(categories)
  → { name: true, sku: true, ... }

anchorSystemColumns(aiOrder)
  → [__check__, ...ai..., actions]

Case study: AI Data Grid Assistant

Natural language filtering for non-technical nursery staff

Ideas were human-led

1. The feature itself
   • Relatively simple to build, plenty of sizzle
   • Foundation for future AI features
2. The output shape
   • Categories, not booleans — after watching AI fail at exact shapes
3. Edge cases caught by human testing
   • Location tree hierarchies, column visibility merge bugs

Implementation was AI-led

1. Plan mode — 15+ times
   • Human sets direction, AI proposes implementation
2. Fed it Vercel's useChat examples
   • Training data cutoff — needed current docs and patterns
3. Fantastic at bulk repetitive work
   • Adding isSearchable to 50+ column definitions across 12 grids
4. Translation layer: pure functions, unit tested
   • mergeAIColumnVisibility() and anchorSystemColumns()

Curate ruthlessly

"The most important word for MCP developers is curate." — Jeremiah Lowin

Summary

1. Humans have ideas and experience
2. LLMs are interns — eager, tireless, occasionally brilliant
3. right context > more (or less) context — web search
4. tools ≠ utils — design for outcomes with agent stories
5. AI output ≠ product shape — build a bridge with unit tests
6. Curate. Be bold