How to Give a Coding Agent a Whole Codebase: The Prompt-Path Pattern

You have a working project on your laptop. A friend asks you to send it over so they can try it. In 2024 you would write a README. In 2026 you write two lines:

"Read every file under ./news-refiner-kit-v1/, install dependencies, set up an isolated environment, and run the smallest working example."

Then you paste this into Claude Code, Cursor, Cline, Aider, or any other coding agent that can read files and run shell. Sixty seconds later the project is running.

That is the entire pattern. This post is about why it works and why it is a better way to ship a codebase than a README.

1. What a coding agent actually needs

Every capable coding agent needs three things to make a codebase run on a new machine:

• A goal. "Get this project to boot the smallest working example."
• A concrete location to read. A directory path, not a link to a blog post.
• Permission to install and execute. Any agent with terminal access already has this; users grant it once when they run the agent.

A README gives the agent a goal but not a directory. A GitHub link gives the agent a directory but no goal. The prompt-path pattern gives both in a single copy-paste.

2. The two-part payload

Every kit on OpenKitMule ships with the same two fields on its detail page:

• Agent prompt. One short paragraph the user copies verbatim. It tells the agent the domain ("Python data-engineering assistant"), the goal ("install Ollama, pull the model, run the first crawl"), and the placeholder {{CODE_PATH}}.
• Code path. A concrete directory like kits/news-refiner-kit-v1/ that appears in the downloaded zip.

The site swaps {{CODE_PATH}} into the prompt at render time, so the user always sees a fully rendered prompt they can paste without editing.

A real example from the News Refiner Kit:

You are a Python data-engineering assistant. Please read all files under
kits/news-refiner-kit-v1/, understand the news crawling pipeline, local
LLM summarizer, and output formats, then help the user install Ollama,
pull the model, and run the first crawl.

3. Why this is better than a README

• READMEs go stale. The code changes; the README does not. An agent reading the current source will always be more accurate than a document that lags.
• READMEs assume one environment. They say "run pip install" but do not know you are on Python 3.13 without pip, or on macOS with a broken Homebrew symlink. An agent inspects and adapts.
• READMEs skip the boring parts. "Set your API key" is one line in a README and 20 minutes of confusion for a first-time user. An agent walks the user through the exact export or dotenv edit.
• Errors get fixed instead of documented. If a dependency install fails, an agent debugs it. A README just told you what should have happened.

4. How to design a good agent prompt

After shipping four kits with this pattern (News Refiner, Agent Scraper, Kronos, PromptForge) we settled on four rules for the setup prompt.

1. State the role in one clause. "You are a Python data-engineering assistant." This narrows the agent's stylistic choices without over-constraining it.
2. Point at a directory, not a file. Agents can list a directory in one tool call. Pointing at one file forces them to guess siblings.
3. Name the runnable goal. "Run the first backtest", "Run the first crawl", "Print token counts for a sample prompt." A verifiable goal is what turns setup from a chat into a task.
4. Warn about the dangerous edge. Trading kits say "verify strategies on paper trading before going live." Data-scraping kits say "respect robots.txt." The prompt is the last piece of copy the agent will read before acting, so put the guardrail there.

5. What agents this works with today

• Claude Code - terminal-native, best-in-class at multi-file edits and long-running sessions.
• Cursor - IDE-based; agent mode reads the workspace and runs commands.
• Cline (open-source) - VS Code extension, bring-your-own-key.
• Aider - Git-first CLI agent, works especially well for "read this repo and make it run."
• ZCode / Kimi Agent / GLM Coding - Chinese-market coding agents with terminal access.
• ChatGPT / Claude web with Advanced Data Analysis or code interpreter - work in a sandbox; the same prompt still applies, just against an uploaded zip.

The pattern is deliberately agent-agnostic. If your setup prompt requires a specific agent, you have written it wrong.

6. What breaks the pattern (avoid these)

• Hardcoded home directories. An absolute user path in a code path means every other user has to edit before pasting. Always use relative paths that match the zip structure.
• Secrets in the repo. If the agent needs a key, the prompt should say "prompt the user for an API key", not assume one exists.
• Massive prompts. A four-page setup prompt tells the agent nothing it could not learn from the code. Keep it under 80 words.
• Missing runnable goal. "Understand the project" is not verifiable. "Run the first crawl and show me the first three items" is.

7. Try it in 60 seconds

Download the News Refiner Kit zip, extract it, then paste the rendered prompt from the product page into Claude Code or Cursor. That is the whole flow. There is no signup, no payment, no marketplace layer.

If you want a smaller, dependency-free example first, try the PromptForge kit - one Python file, stdlib only, useful in about a minute.

8. Related reading

• Original walkthrough of the OpenKitMule workflow (Chinese).
• Estimating LLM API costs before you ship - the sibling article, uses PromptForge as the tool.
• Local LLM news summarization with Ollama - the prompt-path pattern applied to a real Ollama-based pipeline.
• Agent Scraper Kit - a longer example of the prompt-path pattern applied to a scraping framework.
• Kronos - the pattern also works for kits where OpenKitMule does not host the code, only the setup prompt.