If you use ChatGPT, Claude, or any AI agent for code reviews, API design, or writing PRDs, you've probably noticed something: the quality of AI output depends almost entirely on the quality of your context.
The problem is, every time you start a new task, you have to re-explain the rules, conventions, and expected format. It gets repetitive fast.
I've been solving this with something I call FeedPacks — reusable Markdown bundles that package everything an AI needs for a specific task.
The Problem
When I asked Claude to review my PHP code, I'd write something like:
Review this PHP code for best practices, PSR-12 compliance, security issues, and performance problems.
The result was okay, but inconsistent. Claude didn't know my exact standards, what to prioritize, or how to format the output.
The Solution: Structured Context
Instead of a single prompt, I now use a set of Markdown files with a fixed structure:
task-name/
├── instructions.md — AI role, behavior, rules
├── workflow.md — step-by-step process
├── inputs.md — what information to ask for
├── output-format.md — expected output structure
├── examples.md — good and bad examples
├── constraints.md — hard limits
└── checklist.md — quality validation
I feed these files into the AI session before starting the task. The difference is night and day.
Example: PHP Code Review FeedPack
Just the instructions.md file alone gives the AI complete context:
You are a senior PHP developer specializing in code review. Analyze code for PSR-12 compliance, security vulnerabilities, performance issues, and modern PHP 8.x practices. Be constructive. Prioritize by severity. Support each finding with code examples.
Paired with output-format.md, the AI knows exactly how to structure the response — severity rating, explanation, code example, fix suggestion. No more guessing.
Why Not Just Save a Prompt?
A prompt is a one-shot instruction. A FeedPack is a knowledge bundle:
| Aspect | Single Prompt | FeedPack |
|---|---|---|
| Instructions | Yes | Yes |
| Workflow | Maybe | Yes |
| Inputs | No | Yes |
| Output format | Sometimes | Yes |
| Examples | Rarely | Yes |
| Constraints | No | Yes |
| Checklist | No | Yes |
Real Results
Since I switched to this approach:
- Less back-and-forth with the AI
- More consistent output formatting
- Easier to reuse work across sessions
- Simple to share with teammates (it's just Markdown)
Try It Yourself
I've published 10 FeedPacks covering PHP development, Laravel code review, API design, bug triage, SEO audits, PRD creation, and more.
All free, no restrictions: https://github.com/ppronini/aifeedpacks
The format works with ChatGPT, Claude, Cursor, Copilot — any AI that accepts Markdown input.
How to Create Your Own
- Pick a task you do repeatedly with AI
- Create a folder with the 7 .md files listed above
- Fill each one with your specific context
- Next time you need that task done, feed the files to the AI
Start simple. Even just instructions.md + output-format.md will improve your results dramatically.
This approach has saved me hours of re-explaining context to AI tools. Hopefully it helps you too.
Top comments (0)