Is it just me

I've noticed a shift where "modern" coding challenges aren't just about algorithms anymore; they're about how efficiently you can leverage an LLM agent to refactor a legacy codebase or hunt down a race condition. The problem is that the highest-tier capabilities—specifically the massive context windows and the agentic loop in Claude Code—are locked behind a paywall. If you're a junior dev or a student, you're fighting an uphill battle against someone who has a high-token-limit agent doing the heavy lifting in the background.

To get a real sense of how this impacts a workflow, you have to look at the difference between a standard chat interface and a terminal-based agent. When you're using a tool like Claude Code, you aren't just copying and pasting snippets; the agent is reading your file tree, running tests, and iterating on its own mistakes.

If you're trying to mimic this "agentic" power without spending a fortune, you have to get creative with your AI workflow. Here is a basic setup I use to simulate that agentic loop using open-source tools and cheaper API calls:

DIY Agentic Workflow Setup

1. Install an MCP (Model Context Protocol) server. This allows your LLM to actually "see" your local files instead of you manually uploading them.
2. Use a CLI wrapper. Instead of the web UI, use a terminal tool that can pipe shell output back into the prompt.
3. Create a "System Loop" prompt. Use a prompt that forces the AI to think in terms of Plan -> Execute -> Verify.

Here is the prompt structure I use to force a standard LLM to act more like a high-tier agent:

You are an autonomous coding agent. For every request:
1. ANALYZE: Read the provided file context and identify the exact line numbers causing the issue.
2. PLAN: Write a step-by-step plan of the changes required.
3. EXECUTE: Provide the full code block for the modified file.
4. VERIFY: Describe exactly how I should run a test command to verify the fix.
Do not provide conversational filler. Only output the structured blocks.

The productivity gain from this is huge, but it's still a manual process. You're basically acting as the "human router" for the AI, whereas a paid agent handles that routing automatically.

The real-world gap is in the "deep dive" capabilities. A candidate with the Max tier can feed an entire 50-file repository into the context window and ask, "Where is the state management breaking?" and get an answer in seconds. A candidate on a free or lower tier has to spend twenty minutes manually searching for the relevant files to paste into a prompt.

We need to stop treating "knowing how to use AI" as a generic skill and start recognizing that the quality of the tool—and the price of the subscription—directly affects the output. Prompt engineering only gets you so far if the model can't see the rest of the project.