context-engineering-pack

Context Engineering Pack (Summit)

Purpose

This pack integrates the upstream “Agent Skills for Context Engineering” library as a vendored dependency and provides routing rules so agents load only the minimum necessary instructions.

Upstream location (vendored): skills/vendor/agent-skills-context-engineering/skills/

Operating Rules (Summit-specific)

1. Progressive disclosure only

   • Do NOT load every skill into context.
   • Start with a directory scan + short descriptions, then load 1–3 relevant skills’ SKILL.md only when needed.

2. Atomic PR surfaces

   • Prefer changes that touch a single coherent file surface.
   • If multiple surfaces are needed, split into separate PR prompts.

3. Evidence-first for GA

   • When a change affects release/CI/security: include tests, checks, and “how to verify” steps.

Routing Heuristics (when to consult upstream)

Use the upstream library when you see:

• long-running tasks / “agent gets worse over time”
• too much context / tool output spam / “lost in the middle”
• multi-agent coordination issues
• RAG/citations quality problems
• evaluation needs (LLM-as-judge, pairwise evals, bias mitigation)

How to Load Upstream Skills (agent procedure)

1. List candidate skills:

   • Look in: skills/vendor/agent-skills-context-engineering/skills/
   • Identify 3–7 likely skill folders by name.

2. Read only metadata first:

   • For each candidate folder, open the first ~60 lines of SKILL.md (YAML + intro) to get name/description.

3. Select + load:

   • Choose the best 1–3 skills.
   • Load their SKILL.md fully.
   • Apply them to the current task.

4. If still failing:

   • Add at most 1 additional skill.
   • If context is bloated, summarize tool outputs and compress history.

Common “first picks”

• context-fundamentals
• context-degradation
• context-optimization
• multi-agent-architecture
• evaluation / llm-as-judge (when judging outputs)