← Latest brief

Brief #62

41 articles analyzed

Context engineering is escaping vendor frameworks entirely—practitioners are discovering that manual context management fails catastrophically, driving rapid adoption of MCP as infrastructure and custom tooling that treats context as a first-class architectural concern rather than a prompting technique.

Manual Context Files Rot Into Useless Noise

Practitioners maintaining CLAUDE.md files report they degrade into maintenance burdens that actively harm performance. Automated context management (like Claude Code's active recall) solves what discipline alone cannot—context requires tooling support, not human curation.

Stop maintaining manual context files. Invest in automated context management tooling (MCP servers, active recall systems) or use platforms that handle this natively. Test whether your current context preservation strategy degrades over 2+ weeks of use.
@alxfazio: > active recall

Direct practitioner experience: manual CLAUDE.md files become 'rot' and counterproductive. Automated active recall mechanisms solve this where discipline fails.

@iannuttall: who's the smartest context person on here?

Practitioner asking fundamental question about builtin session management vs explicit curation—revealing that 'just maintain context manually' is not a solved problem.

We've added a new command to Claude Code called /insights

Anthropic shipping context reflection as first-class feature validates that raw context accumulation alone is insufficient—needs active synthesis and reflection tooling.


MCP Standardizes Context Integration, Not Just Tools

MCP is being adopted as infrastructure for context management—not merely tool calling. Practitioners are building MCP servers for local data (email archives, project knowledge) to make historical context queryable rather than manually re-explained each session.

Build or integrate MCP servers for your persistent data sources (databases, document repositories, internal APIs) instead of embedding knowledge in prompts. Treat MCP as context infrastructure, not just a tool-calling protocol.
@alexhillman: Oh this rips. Straight onto the NAS!

Practitioner building MCP server (msgvault) to expose 20 years of email as queryable context for Claude, solving data ownership AND context availability simultaneously.

Context-Destroying vs Context-Preserving Tool Adoption

Mandating AI tools that bypass human domain expertise destroys engineer motivation and organizational knowledge. Successful adoption positions AI to amplify existing context (architecture knowledge, codebase mastery) rather than making it redundant.

Before mandating AI tooling, identify what human context (domain knowledge, architectural understanding) must be preserved and amplified. Design adoption to make expertise an accelerator for AI, not redundant to it. Test: does the tool require or eliminate existing knowledge?
Struggling engineers identify with the craft. Thriving engineers identify more...

Direct practitioner report: Claude Code mandate eliminated need for codebase expertise, destroying motivation. Engineers quit when their context becomes worthless overnight.

Custom Context Systems Deliver 93% Task Automation

Building specialized AI systems that own a knowledge domain (project state, client history, decision rationale) enables massive automation of context-heavy judgment work. The bottleneck isn't model capability—it's persistent, domain-specific context architecture.

For context-heavy judgment work (project management, client relations, operational decisions), build custom AI systems that maintain domain-specific state rather than using general-purpose chatbots. Invest in persistent context architecture, not better prompts.
Natalia Quintero runs @every's seven-figure AI consulting...

Practitioner built 'Claudie' that automated 93% of 15hr/week project management by maintaining persistent context across all projects—state, history, decisions, constraints. Intelligence compounds because context doesn't reset.

Real-Time Context Injection Beats Static Retrieval

Skills that automatically inject fresh context from current sources (social, web, forums) before reasoning produce dramatically better outputs than RAG over static knowledge bases. Context freshness is a first-class architectural requirement, not a nice-to-have.

For domains where information freshness matters (trend research, competitive analysis, breaking news), build automated context injection that pulls current data before each reasoning step. Don't rely on RAG over static knowledge bases for time-sensitive work.
I sat down with matt van horn and watched him turn claude code into a...

Practitioner observation: /last30days skill works by automating freshness filtering from multiple sources, injecting current context directly into prompts before reasoning. Prevents building on stale foundations.

Structured Outputs Eliminate Context Negotiation Overhead

Guaranteeing output schema upfront eliminates entire categories of error handling, validation, and retry logic. Schema-as-contract reduces validation code in every downstream consumer—one clear definition cascades through the system.

Define explicit output schemas for AI-generated content instead of parsing freeform text. Use structured output APIs (Anthropic, OpenAI) to guarantee format adherence. Eliminate validation/retry logic by treating schema as contract.
@adocomplete: 28 Days of Claude API - Day 4 - Structured Outputs

Practitioner report: structured outputs eliminated JSON parsing errors and retry logic. Schema itself becomes the context—defining structure upfront eliminates negotiation between expectation and output.

Feedback Loop Velocity Enables Intelligence Compounding

When agents can execute feedback loops automatically and frequently (50x/day vs weekly), compounding returns become visible in days instead of months. The constraint isn't model quality—it's iteration frequency enabling context to accumulate and improve.

Measure and optimize feedback loop frequency in your AI systems. Identify workflows where agents can run experiments/tests/validations automatically at high frequency. Intelligence compounds when iteration happens 10-50x/day, not weekly.
@sqs: It used to take weeks/months to see a net benefit from building better dev fe...

Practitioner observation: agents running feedback loops 50x/day compress timeline from weeks/months to days. Intelligence compounds when feedback preserves across iterations instead of resetting.

Agent Skills as Security Vulnerability Surface

Agent skill definitions that allow embedded code execution create attack vectors when different systems interpret them differently. Context ambiguity (is this documentation, configuration, or executable code?) becomes a security vulnerability, not just a design question.

Audit agent skill definitions for executable code vs documentation. Implement input/output sanitization at boundaries. Understand execution semantics of your agent framework—does it treat skills as markdown or bash scripts? Log all tool executions for security monitoring.
@badlogicgames: Applies to any harness that supports skills plus command execution

Practitioner warning: semantic gap between 'skills as documentation' vs 'skills as executable code' is exploit surface. Different harnesses handle this differently, creating vulnerability when users don't understand execution model.