Adding References & Context Sources to My AI Workflow

Damian Martinez Carmona · September 11, 2025

This is a continuation (Part 2) of my previous post: My New Development Workflow - I Don’t Write Code Anymore. In this article I want to share how I manage references and context sources to help my coding agents (Cursor, Gemini CLI, etc.) stay accurate and up-to-date.

References Folder in ai_docs

Inside my ai_docs/ directory, I keep a subfolder called references/.

  • In references/ I clone repos, documentation, and libraries that I may want to use as extra context.
  • I include instructions in AGENTS.md and in my task templates for when and how the agent should consult those references.
  • Don’t forget to add references/ to .gitignore, because you don’t want to track or commit all that content in Git.

This setup helps when the coding task depends on specifics in a library or external doc. Having those references locally means the agent doesn’t have to guess or rely on possibly stale online info.

Using Context7 MCP Server

Another source of context I rely on heavily is the Context7 MCP server (upstash/context7).

  • It holds up-to­-date documentation for many code libraries.
  • In AGENTS.md and templates I include Context7 as a source of truth whenever tasks involve library usage, API changes, or newer features.
  • This helps ensure the agent has access to the freshest documentation, which reduces errors or outdated assumptions.

Recycling Repos as References

One more thing I often do: I include an existing, well-structured repository as a reference, so I can reuse patterns, structure or boilerplate instead of starting fully from scratch.

For example, I always include my base-agent repo whenever I start a new ADK agent, to build on a basic scaffold of features.

You could do the same with a frontend app boilerplate, API boilerplate, or whatever matches your project domain. The idea is: the agent sees real, working examples in your references so it can borrow useful parts rather than invent them.

Why This Matters

Putting together live references + Context7 + the AGENTS.md + template system means fewer surprises:

  • My coding agents can check real, current docs when needed.
  • Updates to libraries are less likely to break things because the agent has reference material.
  • The workflow stays more stable, less “trial and error” when adding new dependencies or starting new projects.

If you’d like, I can share some examples of my references/ folder, sample templates where I reference Context7, or even a sample AGENTS.md section showing these instructions. Just hit me up via LinkedIn or email.

Twitter, LinkedIn