| name | context-engineering |
| type | workflow |
| description | Strictly enforce context engineering principles to avoid context stuffing, optimize memory architecture, and manage the Research-Plan-Reset-Implement cycle. |
| user-invocable | true |
| allowed-tools | Read, Glob, Grep, Write |
| context | fork |
| agent | technical-director |
| effort | 4 |
| argument-hint | [context source or workflow scope] |
| when_to_use | When prompts, logs, specs, or multi-step work exceed a clean working set and the agent needs to compress, externalize, or reset context before continuing |
1. Overview
Context engineering bridges the gap between static training data and dynamic reality. Context Stuffing (jamming volume without intent) degrades reasoning, increases noise, and leads to hallucinations. Context Engineering treats AI attention as a scarce resource and allocates it deliberately through structure, bounded contexts, and intelligent memory retrieval (RAG / MCP Supermemory). Without this skill, the AI suffers from "Context Hoarding Disorder," leading to goal drift, high latency, and poor execution quality.
2. When to Use
Activate this skill immediately upon detecting the following signs:
- Sign 1: User pastes a massive block of uncurated documents entirely into the context window (e.g., full PRDs, full codebases, thousands of lines of logs).
- Sign 2: The AI's outputs start to become vague, hedged, or inconsistent despite having "all the context", or when the context window is clearly overflowing.
- Sign 3: The user wants to start a multi-step complex workflow spanning many files and iterations.
- Implicit Command: User types
/context or /memory.
3. Strict Process
ULTIMATUM: You are an Agent. You DO NOT have the right to ignore, truncate, or alter the order of these steps, even if you think "the model has a 1 million token context limit anyway."
- [Step 1 - Intent & Boundary Falsification]: Identify exactly what decision the provided context supports. Apply the falsification test: "If I exclude [context element X], what specific failure will occur in [decision Y]?" If there is no concrete failure, the context must be rejected or removed from the active window.
- [Step 2 - Persist vs. Retrieve Classification]: Separate the information. Core constraints and glossary definitions remain in active context. Episodic, project-specific, or historical data must be offloaded and retrieved only when queried. Use
mcp_supermemory_recall for historical lookups instead of keeping them in the prompt.
- [Step 3 - The R-P-R-I Cycle Execution]:
- Research: Gather necessary information.
- Plan: Synthesize findings into a high-density
PLAN.md or SPEC.md.
- Reset: Save crucial lessons to memory using
mcp_supermemory_memory and explicitly ask the user to clear the context window (start a new chat) or summarize everything to drop the past context rot.
