reduce-delegate-framework
Apply R&D framework to optimize prompts and context. Use when optimizing context window usage, reducing prompt size, delegating to specialized agents, or applying systematic context management.
When & Why to Use This Skill
The Reduce & Delegate (R&D) Framework skill provides a systematic methodology for optimizing Claude's context window and prompt efficiency. It enables users to maintain high performance in complex workflows by strategically reducing irrelevant data and delegating subtasks to specialized agents, effectively solving issues like context rot, token overflow, and performance degradation.
Use Cases
- Case 1: Managing long-running conversations where 'context rot' or conflicting information begins to degrade the agent's decision-making quality.
- Case 2: Scaling complex research or development tasks by delegating specific sub-problems to specialized sub-agents with isolated, focused context.
- Case 3: Optimizing token usage and reducing costs by replacing static, bloated memory files with dynamic priming commands and concise output styles.
- Case 4: Improving agent focus during multi-step workflows by clearing 'context pollution' and resetting history through fresh instances while maintaining essential state.
| name | reduce-delegate-framework |
|---|---|
| description | Apply R&D framework to optimize prompts and context. Use when optimizing context window usage, reducing prompt size, delegating to specialized agents, or applying systematic context management. |
| allowed-tools | Read, Grep, Glob |
Reduce & Delegate Framework Skill
Apply the R&D framework to optimize prompts, workflows, and context management.
Purpose
There are only two ways to manage context: Reduce and Delegate. This skill helps you systematically apply both strategies to any context optimization challenge.
When to Use
- Context window approaching limits
- Agent performance degrading over conversation
- Prompts growing unwieldy
- Workflows consuming too many tokens
- Need to scale agent work
The R&D Analysis Process
Step 1: Identify the Context Problem
Categorize the issue:
| Problem Type | Indicator | Primary Strategy |
|---|---|---|
| Context Rot | Old info guiding decisions | Reduce (fresh instance) |
| Context Pollution | Unfocused, tangential | Reduce (remove irrelevant) |
| Toxic Context | Contradictory behavior | Reduce (clear conflicts) |
| Context Overflow | Approaching limits | Delegate (offload work) |
Step 2: Apply Reduce Strategies
For each context element, ask:
- Is this necessary for the current task?
- Can this be loaded on-demand instead?
- Is this information stale or outdated?
- Does this contradict other context?
Reduction techniques:
| Technique | Application |
|---|---|
| Fresh instance | New task type, reset history |
| Output styles | Control verbosity, reduce tokens |
| Focused reads | Specific files vs directories |
| Priming commands | Replace static memory |
| MCP cleanup | Remove unused servers |
Step 3: Apply Delegate Strategies
For complex or parallel work, ask:
- Does this subtask need different context?
- Can this run independently?
- Would a specialized agent perform better?
- Is there parallel work opportunity?
Delegation techniques:
| Technique | Application |
|---|---|
| Sub-agents | Focused tasks with isolated context |
| Background agents | Parallel work, async execution |
| Agent experts | Domain-specific knowledge |
| Spec files | Handoff between agents |
Optimization Workflow
1. Measure current context state
- Use /context command
- Check token consumption
2. Analyze composition
- What's consuming most tokens?
- What's unnecessary?
3. Apply Reduce
- Remove unnecessary context
- Start fresh if needed
- Control output verbosity
4. Apply Delegate
- Offload subtasks
- Use specialized agents
- Enable parallel work
5. Verify improvement
- Measure new state
- Compare performance
Common Optimization Patterns
Pattern: Bloated Memory File
Before:
# CLAUDE.md (5KB+)
Contains: everything about the project
After (Reduce):
# CLAUDE.md (1KB)
Contains: only universals
# .claude/commands/prime.md
Contains: task-specific context loading
Pattern: Long Conversation
Problem: Multi-turn conversation with context rot
Solution (Reduce):
- Start fresh instance
- Use priming command to load current state
- Continue with clean context
Pattern: Complex Research Task
Before:
Primary agent does research -> context polluted
Primary agent implements -> struggles with focus
After (Delegate):
Primary agent delegates research -> sub-agent
Sub-agent returns summary -> primary continues
Primary agent implements -> clean context
Pattern: Parallel Independent Tasks
Before:
Task A -> Task B -> Task C (sequential, context accumulates)
After (Delegate):
Task A (agent 1) \
Task B (agent 2) -> Aggregate results
Task C (agent 3) /
Output Format
When optimizing, report:
{
"analysis": {
"current_state": "Context at 80% capacity",
"primary_issue": "Long conversation with accumulated history",
"secondary_issues": ["Verbose tool outputs", "Unused MCP servers"]
},
"reduce_recommendations": [
{
"action": "Start fresh instance",
"impact": "Reset accumulated history",
"effort": "Low"
},
{
"action": "Apply concise output style",
"impact": "50% reduction in output tokens",
"effort": "Low"
}
],
"delegate_recommendations": [
{
"action": "Create research sub-agent",
"impact": "Isolate research context",
"effort": "Medium"
}
],
"expected_improvement": "40-60% context reduction"
}
Decision Matrix
When to Reduce vs Delegate:
| Situation | Reduce | Delegate |
|---|---|---|
| Stale context | X | |
| Irrelevant context | X | |
| Conflicting context | X | |
| Complex subtask | X | |
| Parallel work | X | |
| Domain expertise needed | X | |
| Context overflow | X | X |
Key Quote
"There are only two ways to manage your context window: Reduce and Delegate. Every technique fits into one or both of these buckets."
Cross-References
- @rd-framework.md - Framework reference
- @context-audit skill - Audit before optimizing
- @context-layers.md - Understanding what to optimize
- @context-rot-vs-pollution.md - Diagnosing the problem
Version History
- v1.0.0 (2025-12-26): Initial release
Last Updated
Date: 2025-12-26 Model: claude-opus-4-5-20251101