tranhieutt/context-engineering

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."

1. [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.
2. [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.
3. [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.
   • Implement: Execute purely based on the dense plan.
4. [Step 4 - Storage & Consolidation]: Upon finishing a milestone, write the generalized knowledge or operational principles into mcp_supermemory_memory.

4. Anti-Rationalizations

These are lazy thoughts that Agents (like you) commonly fall prey to. If an idea in your head matches the Left Column, you MUST immediately obey the Right Column:

| Agent's Lazy Rationalization | Refutation & Mandatory Action | | :--- | :--- | | "I have a massive context window; I can just read all 50 files the user provided without complaining." | Volume != Quality. Context is an attention economy. You MUST explicitly tell the user that "Context Stuffing" dilutes attention. Limit the working context to only the files strictly needed for the immediate decision. | | "I'll just try again and rewrite the code if it hallucinates." | Retries mask bad information architecture. Do not normalize retries. Stop and fix the context structure, break down the problem, or refine the retrieved memory BEFORE writing another line of code. | | "I'll keep all the research logs and failed attempts in the chat history so I remember what I tried." | "Context Rot" kills reasoning. Once a plan is synthesized, the previous dead ends become pure noise. You MUST instruct the user to flush the context or start fresh with a clean SPEC.md. |

5. Verification Gates

You are NOT allowed to end your turn and respond "Context established" or proceed to implementation until the following checks pass:

• [ ] You have successfully filtered out unnecessary broad context and kept only the decision-critical context.
• [ ] You have explicitly called mcp_supermemory_memory to commit important established facts, or mcp_supermemory_recall to fetch past facts instead of asking the user to paste them.
• [ ] You have synthesized the current sprawling context into a dense artifact (e.g. PLAN.md or SPEC.md) and proposed a Context Reset.

6. Red Flags

Immediately STOP and request User intervention if:

• The user demands that you process more than 35k tokens of unstructured, mixed-domain text in a single prompt without allowing you to summarize and discard the noise.
• You detect that the current context window is bloated with >3 failed attempts of the same task. You must halt and enforce a Reset.
• Memory retrieval tools (like mcp_supermemory) fail repeatedly, meaning you are operating blind on historical context.