context-engineer

Installation
SKILL.md

Context Engineer v3.0

Determine what context an agent needs and package it within token budget.

Actions

/context-engineer analyze <task>

Identify what context sources are relevant:

  1. Parse task description for domain keywords
  2. Check for DESIGN.md (frontend tasks)
  3. Check for .spec.md files (spec-driven tasks)
  4. Check memory sources (vault, handoffs, ledgers)
  5. Check codebase files (Glob/Grep for related code)
  6. Output: ranked list of context sources with estimated tokens

/context-engineer load <task>

Assemble context from identified sources:

  1. Run analyze if not already done
  2. Read each source, extract relevant sections
  3. Prioritize by relevance score
  4. Output: assembled context block

/context-engineer prune <budget>

Trim assembled context to fit token budget:

  1. Default budget: 8000 tokens
  2. Remove low-relevance sections first
  3. Summarize long sections instead of including full text
  4. Preserve: interfaces, invariants, design tokens (high signal)
  5. Output: pruned context within budget

/context-engineer inject

Write packaged context to .claude/context-payload.md:

  1. Write pruned context to file
  2. The ralph-subagent-start.sh hook reads this file and injects into teammate prompt
  3. File is gitignored (per Item 0a)

Context Sources (Priority Order)

Source When Relevant Token Cost
.spec.md Task has a spec ~500-2000
DESIGN.md Frontend task ~1000-3000
Related code files Always ~500-5000
Memory (vault) Prior work exists ~200-500
Handoffs Cross-session ~300-800
Vault (future) Curated knowledge ~200-1000

Integration

  • Orchestrator Step 5 (DELEGATE): invoke before spawning teammate
  • ralph-subagent-start.sh: reads .claude/context-payload.md if exists
  • Token budget: default 8000, adjustable per task complexity

Anti-Rationalization

Excuse Rebuttal
"The agent has enough context from the prompt" Agents lose 40% accuracy without structured context. Package it.
"Context engineering is overhead" 5 seconds of packaging saves 5 minutes of wrong output.
"I'll just include everything" Token flooding degrades quality. Prune to signal.
Related skills

More from alfredolopez80/multi-agent-ralph-loop

Installs
1
GitHub Stars
108
First Seen
5 days ago