Context Efficient Tools

Skill: context-efficient-tools

What I do

I prevent large tool results from bloating the context window. When tools return large datasets, I apply filtering, aggregation, and summarisation in code before the result reaches the model. Anthropic found this reduces token usage by up to 98.7% on large MCP tool chains.

When to use me

• When MCP tools might return large datasets (files, search results, database queries)
• When chaining multiple tool calls with large intermediate results
• When bash commands produce verbose output
• When token budget is constrained and tool results are the bottleneck

Core principles

1. Filter before returning — Never pass raw large results to the model
2. Summarise, don't dump — Return counts + samples, not full datasets
3. Store externally, reference internally — Write large results to files, pass the path
4. Progressive disclosure — Start with metadata, drill down only if needed
5. Code does the work — Use bash/scripts to process, not the model

Patterns

Large file reading

# Bad: model sees entire file
cat large_config.json

# Good: extract only what's needed
jq '.database' large_config.json
grep -A5 "relevant_key" large_config.json

Search results

# Bad: 500 matches flood context
grep -r "pattern" .

# Good: count + sample + file list
grep -r "pattern" . | wc -l
grep -r "pattern" . | head -10
grep -rl "pattern" .

Large dataset filtering

# Bad: all 10,000 rows
cat data.csv

# Good: summary + sample
wc -l data.csv && head -5 data.csv
awk -F',' '$3 == "pending"' data.csv | head -10

Storing large outputs

# Store externally, return reference + metadata
some_tool > /tmp/output.txt
echo "Stored $(wc -l < /tmp/output.txt) lines → /tmp/output.txt"
head -5 /tmp/output.txt

Build/install output

# Bad: full verbose output
npm install

# Good: errors and warnings only
npm install 2>&1 | grep -E "error|warn|ERR" | head -20
echo "Exit: $?"

Decision matrix

Anti-patterns to avoid

• ❌ cat on files > 100 lines without filtering
• ❌ Passing full grep output when count + sample suffices
• ❌ Reading entire JSON configs when only one key is needed
• ❌ Letting verbose build output fill context
• ❌ Passing intermediate tool results verbatim to the next tool call

KB Reference

~/vaults/baphled/3. Resources/Knowledge Base/AI Development System/Skills/Session-Knowledge/Context Efficient Tools.md

Related skills

• token-efficiency — Prompt-level efficiency (complements this skill)
• scope-management — Scope determines which tools are called
• parallel-execution — Run independent tool calls simultaneously
• performance — Efficient data processing patterns