Advanced Chunking Patterns Reference

Advanced Chunking Patterns Reference

This reference provides additional chunking strategies and decision frameworks for complex scenarios not covered in the main skill.

Advanced Chunking Strategies

6. Layered Abstraction

Best for: Tasks requiring different levels of detail for different audiences

Example scenario: Creating both executive summary and detailed technical analysis

Approach:

• Session 1: High-level executive summary (research + draft)
• Session 2: Deep technical details (research + draft)
• Session 3: Supporting appendices and data tables

Benefits: Allows prioritization of most important layer first, can stop after any session if time-constrained

7. Hub-and-Spoke

Best for: Central theme with multiple independent supporting analyses

Example scenario: Company overview with deep dives into each business unit

Approach:

• Session 1: Core company profile (central hub)
• Session 2-N: Individual business unit analyses (spokes)
• Final session: Integration and cross-unit insights

Benefits: Hub provides context for all spokes; spokes can be tackled in any order

8. Dependency-Aware Sequencing

Best for: Multi-step processes where later steps depend on earlier decisions

Example scenario: Market entry strategy requiring market research → target selection → go-to-market plan

Approach:

• Explicitly map dependencies upfront
• Complete foundational analysis first
• Use findings to inform subsequent sessions
• Document decisions between sessions

Benefits: Ensures quality by building on validated foundations

9. Spiral Development

Best for: Complex creative projects benefiting from multiple refinement passes

Example scenario: Building a comprehensive training curriculum

Approach:

• Pass 1: Complete curriculum structure at high level (all modules outlined)
• Pass 2: Develop 50% of content in depth
• Pass 3: Develop remaining 50% in depth
• Pass 4: Quality pass and polish

Benefits: Maintains holistic view while making steady progress; can adjust scope mid-project

10. Resource Preloading

Best for: Tasks requiring heavy context that will be referenced throughout

Example scenario: Analyzing code changes across a large codebase

Approach:

• Session 0: Load and understand architecture/key files (no deliverable)
• Session 1-N: Actual analysis work with preloaded context
• Optimization: Summarize key context to carry forward efficiently

Benefits: Amortizes context loading cost across multiple work sessions

Decision Framework for Strategy Selection

Task Classification Matrix

Hybrid Approaches

Real-world tasks often benefit from combining strategies:

Example: Quarterly business review across 5 departments

Hybrid approach:

1. Sequential (by quarter) + Parallel Track (by department)
2. Structure: Q1 Dept A-B, Q1 Dept C-E, Q2 Dept A-B, Q2 Dept C-E, etc.

Example: Product comparison with deep research

Hybrid approach:

1. Depth Progression (outline → content) + Parallel Track (by product)
2. Structure: Outline all products, develop Product A deep, develop Product B deep, synthesize

Estimation Refinement Techniques

Interactive Scoping

When task size is unclear, use a scoping conversation:

"Before I start, let me understand the scope to plan this well:
- Depth: Are you looking for [surface-level summary] or [detailed analysis with citations]?
- Breadth: Should I cover [subset] or [comprehensive coverage]?
- Output: Do you need [bullet points] or [formatted report/presentation]?

Your answers help me estimate whether we should chunk this task."

Token Checkpoints

For borderline tasks, establish checkpoints:

"This might be achievable in one session, but it's close. I'll:
1. Complete the first major section
2. Check token usage at that point
3. Decide whether to continue or split remaining work

Sound good?"

Progressive Disclosure with User

Don't present all options at once:

"This will exceed our token budget. I see two promising approaches:
A. Split by time period (quarters)
B. Split by analysis type (quantitative vs qualitative)

Which makes more sense for your needs?"

Chunking Anti-Patterns

Avoid these common mistakes:

❌ Arbitrary Splits

Bad: "Let's do half now and half later" Why it fails: No logical boundary, hard to resume coherently Better: Identify natural boundaries (themes, time periods, categories)

❌ Unequal Chunks

Bad: Session 1 is 80% of work, Session 2 is 20% Why it fails: Defeats purpose of chunking; Session 1 still exceeds budget Better: Balance work across sessions to ~60-70% of budget each

❌ Chunk Interdependence

Bad: Session 2 requires re-analyzing everything from Session 1 Why it fails: Duplicates work, compounds token costs Better: Design chunks to be relatively self-contained

❌ No Synthesis Plan

Bad: Create 5 independent analyses with no plan to integrate Why it fails: User left to connect dots; misses overarching insights Better: Always include final integration/synthesis session

❌ Over-Chunking

Bad: Split a 70% budget task into 4 mini-sessions Why it fails: Overhead of context switching; destroys flow Better: Only chunk when necessary (80%+ budget risk)

Context Handoff Techniques

Effective Session Summaries

At end of each chunk, provide:

1. What was completed: Specific deliverables and findings
2. Key insights: 3-5 critical points to carry forward
3. Next session focus: Clear starting point
4. Open questions: Unresolved items to address

Example:

Session 1 complete. We analyzed Q1-Q2 revenue data and found:
- 15% growth in Q2 driven by enterprise segment
- Consumer segment flat YoY
- Churn increased 3% in May (investigate further)

Next session: Analyze Q3-Q4 with focus on:
- Did enterprise growth continue?
- What happened with consumer rebound efforts?
- Did churn stabilize?

User-Driven Handoffs

Coach users on what to say when resuming:

"When you're ready for Part 2, start a new conversation and say:
'Continue the Q3-Q4 analysis. In Part 1, we found [X, Y, Z]. Here are the Q3-Q4 files: [upload]'

This gives me the context I need to pick up smoothly."

Special Considerations

Working with External Tools

When chunking involves external tool research:

• Front-load research: Do heavy web searches early to inform all sessions
• Document sources: Save URLs/citations to reference across sessions
• Cache findings: Summarize research to avoid re-searching same topics

Artifact-Heavy Tasks

When creating large artifacts (decks, documents):

• Template first: Establish structure/template in Session 1
• Parallel development: Can work on different sections independently
• Final polish separate: Reserve last session for formatting/refinement

Real-Time Collaboration

If user wants to work synchronously through chunks:

• Time-box sessions: "Let's spend ~15 min on each section"
• Quick transitions: Minimize handoff overhead between chunks
• Live pivots: Be ready to adjust plan based on findings

Measuring Success

Good chunking strategies should:

✅ Complete within token budget for each session ✅ Produce valuable interim deliverables (not just work-in-progress) ✅ Minimize redundant work across sessions ✅ Feel natural to resume (clear continuity) ✅ Allow flexibility to reprioritize between sessions

Poor chunking feels like:

❌ Constantly running out of tokens mid-session ❌ Extensive recap needed to resume work ❌ Deliverables only make sense when all chunks complete ❌ Frequent context re-loading of same information