Context Engineering Advisor
Diagnose whether you're doing context stuffing (volume without intent) or context engineering (structure for attention). Build memory architecture, define context boundaries, and implement the Research → Plan → Reset → Implement cycle.
You pasted the entire PRD, all the user interviews, and last quarter's meeting notes into Claude — and the response is still vague and hedged. That's not a prompt problem. It's an architecture problem: volume ≠ quality, and context windows aren't free.
Who it's for: PMs whose AI outputs feel mediocre despite stuffing huge context, AI engineers designing agent workflows, product leaders building AI-powered features, anyone burning tokens without seeing accuracy gains, teams with agent chains that pass everything downstream
Example
"My AI outputs are mediocre even though I'm giving it lots of information — diagnose what's wrong" → Run 5 diagnostic questions, produce a Context Manifest (persist vs. retrieve vs. exclude), design two-layer memory architecture, and apply Research→Plan→Reset→Implement to eliminate context rot
CLAUDE.md Template
New here? 3-minute setup guide → | Already set up? Copy the template below.
npx ccplaybooks add context-engineering-advisor# Context Engineering Advisor
Guide product managers through diagnosing whether they're doing **context stuffing** (jamming volume without intent) or **context engineering** (shaping structure for attention). Identify context boundaries, fix "Context Hoarding Disorder," and implement tactical practices like bounded domains, episodic retrieval, and the Research→Plan→Reset→Implement cycle.
**Key Distinction:** Context stuffing assumes volume = quality ("paste the entire PRD"). Context engineering treats AI attention as a scarce resource and allocates it deliberately.
## Context Stuffing vs. Context Engineering
| Dimension | Context Stuffing | Context Engineering |
|-----------|------------------|---------------------|
| **Mindset** | Volume = quality | Structure = quality |
| **Approach** | "Add everything just in case" | "What decision am I making?" |Get new playbooks like this one
One email a week with new Claude Code workflows. Free, like everything here.
No spam. Unsubscribe anytime.
What This Does
Diagnoses whether your AI workflows are doing "context stuffing" (jamming volume without intent) or "context engineering" (shaping structure for attention). Identifies context boundaries, builds a two-layer memory architecture, and implements the Research→Plan→Reset→Implement cycle to prevent context rot.
Not prompt engineering — information architecture design.
Quick Start
mkdir -p ~/Documents/ContextEngineering
mv ~/Downloads/CLAUDE.md ~/Documents/ContextEngineering/
cd ~/Documents/ContextEngineering
claude
Describe your current AI workflow and symptoms. Claude walks you through the 5 diagnostic questions and produces a Context Manifest + action plan.
The 5 Diagnostic Questions
- What specific decision does this support?
- Can retrieval replace persistence?
- Who owns the context boundary?
- What fails if we exclude this?
- Are we fixing structure or avoiding it?
If you can't answer #1 or #4, delete the context.
The Research → Plan → Reset → Implement Cycle
- Research — chaotic context allowed, gather broadly
- Plan — synthesize into a high-density SPEC.md
- Reset — clear the entire context window
- Implement — fresh session using only the plan
This is the fix for context rot. Skipping the Reset is why agents "lose the thread" on long tasks.
Five Markers of Context Stuffing
- Reflexively expanding context windows
- Persisting everything "just in case"
- Chaining agents without boundaries
- Adding evaluations to mask inconsistency
- Normalized retries ("it works if you run it 3 times")
Accuracy drops below 20% past ~32k tokens. Using RAG with 25% of available tokens preserves 95% accuracy.
Tips & Best Practices
- Treat tokens as scarce. "Infinite context" marketing doesn't match engineering reality.
- Retrieve with intent, not completeness. If information is referenced in <20% of interactions, retrieve it; don't persist it.
- Assign a context boundary owner. Without one, context grows indefinitely.
- Run quarterly context audits. Re-apply the 5 diagnostic questions.
- If you're retrying prompts 3+ times, don't tune the prompt — fix the structure.
Common Pitfalls
- Believing more tokens = better outputs (they don't past ~32k)
- Normalizing retries instead of restructuring
- Passing full context between chained agents
- Mixing always-needed and episodic context in the same layer
- Skipping the Reset phase and letting context rot poison implementation