agentleFS
mikeparcewskimikeparcewski/wicked-garden/skills/agentic/context-engineering/SKILL.md

wicked-garden-agentic-context-engineering

Skill9 starsChanged 7 months agoMITView on GitHub

Context window management, token optimization, and memory patterns for efficient multi-agent systems.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 4 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Only load relevant context based on the current task.

    line 45
  2. Distinctive:

    Use JSON/structured formats instead of prose to reduce tokens.

    line 48
  3. Distinctive:

    Before: "The user's name is John Smith..." (verbose)

    line 51
  4. Distinctive:

    Lazy load details: Only when needed

    line 133

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Context Engineering

Techniques for managing context windows, optimizing token usage, and designing efficient memory systems for agentic applications.

Context Window Fundamentals

Context Window: Maximum tokens an LLM can process in a single request (input + output).

Limits vary by provider and model. Check the active model's documentation for the exact value.

Token Efficiency Matters:

  • Cost: Charged per token (input + output)
  • Latency: More tokens = slower response
  • Quality: Irrelevant context can confuse model

State Management Patterns

PatternUse whenProsCons
SharedAgents need synchronized viewConsistency, simple coordinationContention, single point of failure
IsolatedAgents operate independentlyNo contention, parallel executionInconsistency possible, harder to coordinate
CheckpointedLong-running processes, need recoveryFault tolerance, replayabilityStorage overhead, consistency complexity

Token Optimization Techniques

1. Aggressive Summarization

Compress old context into summaries to reduce token usage.

2. Selective Context Loading

Only load relevant context based on the current task.

3. Structured Compression

Use JSON/structured formats instead of prose to reduce tokens.

Example:

  • Before: "The user's name is John Smith..." (verbose)
  • After: {"name": "John Smith", ...} (compact)

4. Lazy Loading

Load details only when explicitly needed.

5. Reference Instead of Embedding

Reference external documents instead of embedding full text.

See refs/selective-loading.md and refs/caching-and-optimization.md for code examples and detailed strategies.

Memory Patterns

MemoryScopeSizeRetention
Short-term (working)Current session/task1K-10K tokensMinutes to hours
Long-termCross-session, permanentUnbounded (vector DB)Days to forever
EpisodicHistorical eventsSummaries storedVaries by importance

See refs/compression-techniques.md for implementation patterns.

Prompt Engineering for Agents

Role Definition

Be specific about agent's role and boundaries.

Example:

You are a Python code reviewer specializing in security.
Your job is to identify security vulnerabilities.
You do NOT review style or performance.

Task Specification

Clear, actionable instructions with explicit format.

Bad: "Review this code." Good: "Review for security: 1) SQL injection 2) Input validation 3) Secrets. Output: JSON with vulnerabilities."

Format Control

Specify exact output format to reduce tokens.

Few-Shot Examples

Show examples for complex tasks.

See refs/selective-loading.md for detailed prompting patterns.

Context Loading Strategies

StrategyProsCons
AnticipatoryFaster response time (load before needed)May load unnecessary data
Just-in-Time (JIT)Minimal token usage (load only when needed)Latency on each request
HybridBalanced (core context + JIT for task-specific)More complex implementation

Cost Modeling

Token Cost Calculation

Track input and output tokens separately. Rates vary by model (typically $0.003-0.075 per 1K tokens).

Budget Enforcement

Set hard token limits per agent/session to prevent runaway costs.

Multi-Agent Cost Attribution

Track costs per agent to identify expensive components.

See refs/cost-calculation-budget.md and refs/cost-optimization-reporting.md for detailed cost strategies.

Context Window Strategies by Agent Pattern

Sequential Pattern: Pass only output of previous agent, not entire chain.

Hierarchical Pattern: Parent gets summaries from children, children get only relevant task context.

Collaborative Pattern: Shared context (compressed), each agent adds only delta.

Autonomous Pattern: Minimal shared context, isolated context per agent.

Quick Wins

  1. Compress old messages: Summarize history > 20 messages
  2. Use structured outputs: JSON instead of prose
  3. Lazy load details: Only when needed
  4. Set token budgets: Hard limits per agent/session
  5. Monitor token usage: Track and optimize high-cost agents

References

  • refs/compression-techniques.md - Conversation summarization, deduplication, entity compression
  • refs/selective-loading.md - Relevance filtering, time decay, token-budgeted retrieval
  • refs/caching-and-optimization.md - Prompt caching, semantic caching, batching, cost-aware model selection
  • refs/cost-calculation-budget.md - Token pricing, cost calculation, budget management
  • refs/cost-optimization-reporting.md - Cost estimation, optimization strategies, reporting

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.