$ git log --reverse context/

Context Engineering Timeline

Three years of research into how AI agents should understand codebases, maintain memory, and learn team-specific patterns. From .context.md files to the Agent Skills standard.

This is the story of moving from prompt gambling to engineering discipline—one standard at a time.

The Evolution

2020-2023

Early Experiments: .context Files

Before standards existed, developers manually created .context.md files in their repos. Simple markdown files explaining architecture, conventions, and patterns. The AI would read them... if you remembered to tell it to look.

# .context.md
We use ECS for deployments
Test with Vitest, not Jest
API routes are kebab-case

The problem: Static, manual, no discovery, no standards.

2024

Codebase Context Specification (CCS)

I formalized the .context pattern into a proper specification at codebasecontext.org. One of the first standards for AI codebase context. Introduced hierarchical context (project, directory, file levels), multiple formats (Markdown, YAML, JSON), and conventions for AI consumption.

The breakthrough: Standardized format, early tool adoption.
The limitation: Still static. No dynamic loading.

2024-2025

mem8: Workspace Memory Research

Built mem8 as a Claude Code plugin to research dynamic context and workspace memory. Questions explored: How should agents maintain context across sessions? How do you surface the right information at the right time? What's the difference between project context and team knowledge?

The insight: Context isn't just documentation—it's about when to load information, what the agent needs for each task, and how to keep knowledge updated.

Dec 2025

Agent Skills: The Open Standard

Anthropic publishes Agent Skills as an open standard. Solves everything we'd identified: dynamic discovery, task-based loading, cross-platform portability (Claude, Cursor, VS Code), git control, and team ownership. Microsoft and OpenAI immediately adopt compatible formats.

The realization: This is the standard. The evolution is complete.

2025+

Production Implementation

Now specializing in Agent Skills implementation for startups. Building foundry as a curated Skills registry plus consulting services for custom implementations. The research phase is over—this is production work.

View Solutions →

Key Concepts

Static vs Dynamic Context

Static (CCS era): .context.md files sit in repo. AI reads them if told to.

Dynamic (Skills era): Agent discovers and loads Skills based on current task. No manual prompting required.

Documentation vs Procedural Knowledge

Documentation: "We use ECS for deployments"

Procedural: "Run this exact command with these flags, check these logs, rollback if X happens"

Prompts vs Skills

Prompts: "Remember we use Fargate with..." (context evaporates)

Skills: Git-controlled knowledge that agents load automatically (ground truth)

Individual vs Team Context

Individual: Each dev builds their own context through conversation

Team: Shared Skills library ensures everyone's AI follows same patterns

> Implement Agent Skills

Stop Gambling. Start Engineering.

Three years of context research, now available as production-ready Skills implementation for your startup. Let's build your Skills library.