Level 8Lesson 0⏱️ 120 min

Level 8 Capstone

Design a real context pipeline - decide what the model sees, in what order, and what gets thrown away - with an interactive budget planner

Engineer the Desk, Not Just the Prompt

Over eight lessons you learned to treat context as a finite resource and design the four pillars - instructions, retrieval, memory, and tools - into one tight, high-signal context window. This capstone has you do it for a real agent task, end to end.

No code required. The deliverable is a context plan: a clear, defensible design for what your agent sees on every turn, why, and what it prunes.

Your Mission

Pick an agent task you care about, then produce a complete context design with five parts:

1. THE TASK & BUDGET - what the agent does + a token budget split
                       across the four pillars (use the planner below)

2. INSTRUCTIONS      - a system prompt at the "right altitude":
                       clear heuristics, not brittle rules or vague fluff

3. RETRIEVAL         - what to fetch, when (up-front / just-in-time /
                       hybrid), and how you keep it fresh

4. MEMORY            - how the agent stays coherent over a long task
                       (compaction, notes, or sub-agents)

5. PRUNING & FAILURES- your re-ranking + ordering plan, and which
                       context failure you're most at risk of + the fix

🏆 Capstone Project Helper

Use the Context Budget Planner to build parts 1, 3, and the spine of your design. Set a total budget, split it across the four pillars, and watch the "desk" fill - it warns you when you blow the budget, starve the model of headroom, or let one pillar hog the window. Answer the four questions and it generates a copy-paste context-plan.md.

🧮 Context Budget Planner

Allocate a token budget across the four pillars, watch the desk fill up, and generate a context plan you can copy out.

Agent taskTotal context budget (tokens)

Instructions

Retrieval

Memory

Tools

The desk (50,000 tokens)28,000 used · 22,000 headroom

InstructionsRetrievalMemoryToolsHeadroom

Name your task to get a verdict.

When do you fetch?What do you fetch?How do you compress?When do you discard?

context-plan.md

# Context Plan - (your task)

## Budget (50,000 tokens)
- Instructions : 4,000
- Retrieval    : 16,000
- Memory       : 6,000
- Tools        : 2,000
- Headroom     : 22,000   (room for the model to reason + reply)

## The four questions
1. WHAT to fetch    : (which docs / files / facts?)
2. WHEN to fetch    : Hybrid (rules up front, files just-in-time)
3. HOW to compress  : (summarize? truncate? re-rank to top-k?)
4. WHEN to discard  : (evict stale or finished context when?)

## Pruning checklist
- [ ] Re-rank retrieved candidates down to a tight top-k
- [ ] Highest-signal material placed at the top or bottom (avoid the middle)
- [ ] Tool set is minimal and non-overlapping
- [ ] Old tool results / finished context cleared each turn
- [ ] Token budget tracked per task, with an alert on overrun

Read the warnings as lessons: if the planner says "Retrieval is over 40% of the budget," that's the exact instinct (dump everything) the whole level warns against. A balanced desk with real headroom is the goal.

Finish the Other Parts

Part 2 - Instructions (from Lesson 67)

Write the actual system prompt at the right altitude. Include role, key heuristics, output format, and one canonical example. Avoid the brittle-rules and vague-fluff traps.

Part 4 - Memory (from Lesson 69)

State how your agent survives a long task: compaction, a NOTES.md, sub-agents - or a mix - and what must never be lost in a summary.

Part 5 - Pruning & failures (from Lessons 71-72)

Describe your re-rank + ordering plan, then name the one context failure you're most exposed to (overload? stale? lost-in-the-middle?) and your defense.

What "Great" Looks Like

A strong submission:

Treats context as a budget - every pillar earns its tokens, with real headroom left for the model to think.
Chooses retrieval timing deliberately and has a freshness answer.
Has a concrete memory plan for long tasks, not "the model will remember."
Keeps tools few, clear, and non-overlapping.
Names its biggest context-failure risk by symptom and gives the matching fix.

What You Have Learned in Level 8

L65 - From wording to wiring; context is everything the model sees
L66 - The attention budget; context rot and lost-in-the-middle
L67 - Pillar 1: instructions at the right altitude
L68 - Pillar 2: retrieval, up-front vs just-in-time
L69 - Pillar 3: memory, compaction, notes, sub-agents
L70 - Pillar 4: few, clear, non-overlapping tools
L71 - Assembling context: the pipeline, re-ranking, token budgets
L72 - When context goes wrong: the failure list and fixes

You can now see the context window - and engineer it. When an agent misbehaves, you'll reach past "use a bigger model" and ask the better question: what does it have on its desk right now, and is it the smallest high-signal set that does the job?That question is the whole discipline, and it's yours now.

🧠 Check your understanding

5 quick questions on Level 8. Answer each, then check your score.

1. Context engineering differs from prompting because it...

2. The "attention budget" / context rot means...

3. What are the four pillars of context?

4. "Lost in the middle" refers to...

5. The retrieval pillar is about...

💬

How was Level 8?

Your feedback helps improve this course. Takes 30 seconds.

← When Context Goes Wrong