Context Synthesis

Efficient multi-source context gathering that minimizes token usage while maximizing relevant information.

When to Use

• Starting stakeholder discovery/interviews
• Researching new features or domains
• Building context for analysis tasks
• Synthesizing information from multiple sources

Core Principle

Gather silently, synthesize briefly, share relevantly.

Token efficiency comes from:

1. Parallel MCP tool calls (not sequential)
2. Filtering irrelevant results before presenting
3. Structured summaries over raw dumps

---

Context Gathering Pattern

Step 1: Parallel Information Retrieval

Execute these in parallel (single tool call block):

# All four in parallel - not sequential
mcp__plugin_claude-mem_mem-search__search(query="{keyword}")
mcp__serena__list_memories()
Glob(pattern="**/features/*_FEATURE.md")
WebSearch(query="{domain} best practices 2025")

Step 2: Selective Deep Reads

Based on Step 1 results, read only high-relevance items:

# Only if memory mentions relevant topic
mcp__serena__read_memory(memory_file_name="relevant_memory")

# Only if glob found matching specs
Read(file_path="/path/to/relevant/*_FEATURE.md")

# Only if search returned actionable results
WebFetch(url="most_relevant_url", prompt="extract specific info")

Step 3: Structured Synthesis

Present findings in structured format:

**Context Summary** ({feature/topic})

| Source | Key Finding | Relevance |
|--------|-------------|-----------|
| Memory | Past decision X | Direct |
| Spec FEATURE_A | Similar pattern Y | Reference |
| Web | Industry trend Z | Background |

**Implications for Current Task:**
- [Key implication 1]
- [Key implication 2]

---

Source Priority Order

Priority	Source	When to Use	Token Cost
1	claude-mem	Always first	Low
2	serena memories	Project context	Low
3	Existing specs	Pattern reference	Medium
4	WebSearch	Industry context	Medium
5	WebFetch	Deep dive needed	High

---

Anti-Patterns

Anti-Pattern	Problem	Better Approach
Sequential tool calls	Slow, inefficient	Parallel execution
Reading all files	Token waste	Selective deep reads
Dumping raw results	Cognitive overload	Structured synthesis
Skipping memory check	Miss past decisions	Always check first
WebFetch everything	High token cost	Only for high-value URLs

---

Integration with Other Skills

With requirements-discovery

1. context-synthesis gathers background
2. requirements-discovery conducts interview
3. Context informs question prioritization

With architecture

1. context-synthesis gathers existing patterns
2. architecture analyzes against patterns
3. Context validates decisions

---

Quick Reference

# Minimal context check (fast)
mcp__plugin_claude-mem_mem-search__search(query="{topic}")
mcp__serena__list_memories()

# Standard context gathering (balanced)
# Add: Glob for existing specs, WebSearch for trends

# Deep context research (comprehensive)
# Add: WebFetch for detailed sources, multiple memory reads