Why I Deleted 80% of My AI System Prompt (And Got Twice the Results)

Published: 2026-10-25 · 6 min read · Poorvith M P

Why I Deleted 80% of My AI System Prompt (And Got Twice the Results) Published: October 25, 2026 Author: Poorvith M P Platform Link: [Read on X](https://x.com/poorvith_mp/status/2114159405310804424) ![Why I Deleted 80% of My AI System Prompt (And Got Twice the Results) Cover](/blog-assets/why-i-deleted-80-percent-ai-system-prompt/cover.png) Six months ago, my AI system prompt looked like a hostage negotiation letter. It was sixty-four lines of pure, unfiltered desperation. I had rules written in ALL-CAPS, three long paragraphs explaining how I like my variable names formatted, a grumpy lecture on which JavaScript libraries I hate, and four separate reminders begging the model to "please think step by step before touching any code." Have you ever sent a friend to the grocery store with ten verbal items, zero written list, and a complicated instruction on which brand of milk to buy? Ten minutes later, they walk back into your kitchen holding a bag of tortilla chips and a pack of AA batteries. That is your AI coding agent on turn four of a bloated system prompt. The more rules you cram into your system prompt, the faster your AI forgets all of them. ### The 3-Turn Amnesia Problem Here is what actually happens inside an AI session when you hand it a 50-line wish list of instructions at the start of a chat. Turn one feels like magic. The model follows your formatting rules, speaks politely, and checks the right folders. Turn two is decent. By turn three, after the agent reads four code files and hits one messy terminal error, your carefully crafted 50-line prompt gets buried under thousands of words of fresh code and error logs. By turn five, the agent has complete amnesia. It is happily breaking Rule #2 from your prompt while writing you a three-paragraph apology for breaking Rule #4. Earlier this week, I made a joke on X about how developers treat `.md` instruction files like corporate executives: [https://x.com/poorvith_mp/status/2112170731387973752](https://x.com/poorvith_mp/status/2112170731387973752) We laugh about calling `.md` a "Managing Director," but expecting a single plain-text file to manage your entire codebase is the exact reason our agents keep losing their minds mid-task. I dug into the math behind why giant 1-million-token context windows make coding models slower and dumber in my earlier breakdown on Context Rot: [https://x.com/poorvith_mp/status/2103707567520374796](https://x.com/poorvith_mp/status/2103707567520374796) What happens when your prompt is too long: - Attention dilution: The model spreads its focus across 50 rules that have nothing to do with the bug you are fixing right now. - Token tax: You pay for those 60 lines of text on every single turn of the conversation, even when you just ask a 5-word question. - Rule collisions: When two long rules slightly contradict each other, the AI guesses randomly instead of asking you. ### Flipping the Pyramid (How Top Builders Structure Agents) A few friends of mine who write some of the sharpest essays on X have a rule for writing: if you try to emphasize every sentence on a page with bold text, nothing is emphasized. Yesterday, I posted a quick thought about why developer prompts fail after turn three, and a builder named Martin (@martin_ict_algo) replied with a diagram that every developer using AI should print out and tape to their monitor: [https://x.com/poorvith_mp/status/2113502678747644037](https://x.com/poorvith_mp/status/2113502678747644037) [https://x.com/martin_ict_algo/status/2113528945799008668](https://x.com/martin_ict_algo/status/2113528945799008668) Martin pointed to the "Trust Pyramid" (originally shared by engineer Lauren Tan). Look at the order of priority from most reliable to least reliable: 1. Code structure and strict types first. 2. Linters, automated tests, and CI checks second. 3. Small, on-demand skill files third. 4. English prose rules in the prompt dead last. Almost every developer I know builds this pyramid completely upside down. We write loose, untyped code with zero automated checks, and then we try to save ourselves by writing a 500-word English essay in `CLAUDE.md` or `.cursorrules` begging the AI to write clean code. English prompts are polite suggestions; compilers and type checks are the law. If you are wasting five lines of your system prompt explaining what fields belong inside a user object, delete those five lines immediately. Put a strict TypeScript interface or Python Pydantic schema in your code instead. When the AI makes a mistake, the compiler slaps its hand in 0.2 seconds with a red error line, and the AI fixes it on the spot. ### The "How" vs. The "Never" Rule When I decided to clean up my own setup last month, I sat down and deleted 80% of my global rules file. To decide what stayed and what got deleted, I split every instruction into two buckets: - Bucket 1: The "Never" Rules (Always-On). These are the landmines that will ruin your week if the agent steps on them even once. - Never commit a live `.env` file or secret key. - Never run `git push --force` on the main branch. - Never delete a database migration file without asking me first. - Bucket 2: The "How" Rules (On-Demand Skills). These are step-by-step recipes for specific jobs you only do occasionally. - How to write a new database migration. - How to run a security audit on open ports. - How to format a release note. Your main system prompt (`AGENTS.md` or `CLAUDE.md`) should only contain Bucket 1. Keep it under 15 lines of blunt, unmistakable "Never" rules. Bucket 2 gets kicked out of the main prompt completely. Instead, put each recipe into its own small, separate file inside a local `skills/` folder. (This exact headache is why I built and open-sourced Skillary, a collection of 315 small, modular skill files that agents only open when they need them). When my agent needs to audit a database migration today, it doesn't carry database rules around in its head all morning. It opens the database skill file for thirty seconds, follows the checklist, and closes the file. Here is the Before vs. After from my own daily coding sessions after cutting 80% of my prompt: - Token usage per session: Dropped by 62% overnight because the agent stopped re-reading 50 lines of irrelevant instructions on every message. - Response speed: First-token replies feel almost instant because the base system prompt is tiny and stays cached. - Rule compliance: Because there are only a handful of hard boundaries left in the main file, the agent actually obeys them past turn ten. ### Key Takeaways: Your 3-Minute Prompt Purge Today Earlier this month, four major AI models launched within twenty days of each other, and I wrote about why I didn't switch my setup for any of them: [https://x.com/poorvith_mp/status/2106628505358533039](https://x.com/poorvith_mp/status/2106628505358533039) A brand-new frontier model won't save a cluttered workflow. If you want your current AI agent to feel twice as smart today without spending an extra dollar, open your `CLAUDE.md`, `.cursorrules`, or `AGENTS.md` file right now and do this three-step purge: 1. Delete every rule a tool can enforce. If ESLint, Prettier, Ruff, or TypeScript can catch the mistake automatically, delete the English sentence about it from your prompt. 2. Evict your "How-To" recipes into separate files. Move deployment checklists, testing guides, and styling rules into small markdown files inside a `skills/` folder so they only load when needed. 3. Cut your main file down to 10 hard "Never" laws. Short sentences. Zero fluff. Only the boundaries that protect your code and secrets from disaster. Question for you: Go open your main AI rules file (`CLAUDE.md` or `.cursorrules`) right now. How many lines long is it today, and what is the one rule in there that your AI ignores the most? Drop your line count in the replies!

Why I Deleted 80% of My AI System Prompt (And Got Twice the Results) | Poorvith M PHow cutting a 64-line system prompt down to 10 hard boundaries and moving task recipes into on-demand Skillary files dropped token usage by 62% and cured 3-turn amnesia.

Why I Deleted 80% of My AI System Prompt (And Got Twice the Results)

Published: 2026-10-25 · 6 min read · Poorvith M P

Why I Deleted 80% of My AI System Prompt (And Got Twice the Results) Published: October 25, 2026 Author: Poorvith M P Platform Link: [Read on X](https://x.com/poorvith_mp/status/2114159405310804424) ![Why I Deleted 80% of My AI System Prompt (And Got Twice the Results) Cover](/blog-assets/why-i-deleted-80-percent-ai-system-prompt/cover.png) Six months ago, my AI system prompt looked like a hostage negotiation letter. It was sixty-four lines of pure, unfiltered desperation. I had rules written in ALL-CAPS, three long paragraphs explaining how I like my variable names formatted, a grumpy lecture on which JavaScript libraries I hate, and four separate reminders begging the model to "please think step by step before touching any code." Have you ever sent a friend to the grocery store with ten verbal items, zero written list, and a complicated instruction on which brand of milk to buy? Ten minutes later, they walk back into your kitchen holding a bag of tortilla chips and a pack of AA batteries. That is your AI coding agent on turn four of a bloated system prompt. The more rules you cram into your system prompt, the faster your AI forgets all of them. ### The 3-Turn Amnesia Problem Here is what actually happens inside an AI session when you hand it a 50-line wish list of instructions at the start of a chat. Turn one feels like magic. The model follows your formatting rules, speaks politely, and checks the right folders. Turn two is decent. By turn three, after the agent reads four code files and hits one messy terminal error, your carefully crafted 50-line prompt gets buried under thousands of words of fresh code and error logs. By turn five, the agent has complete amnesia. It is happily breaking Rule #2 from your prompt while writing you a three-paragraph apology for breaking Rule #4. Earlier this week, I made a joke on X about how developers treat `.md` instruction files like corporate executives: [https://x.com/poorvith_mp/status/2112170731387973752](https://x.com/poorvith_mp/status/2112170731387973752) We laugh about calling `.md` a "Managing Director," but expecting a single plain-text file to manage your entire codebase is the exact reason our agents keep losing their minds mid-task. I dug into the math behind why giant 1-million-token context windows make coding models slower and dumber in my earlier breakdown on Context Rot: [https://x.com/poorvith_mp/status/2103707567520374796](https://x.com/poorvith_mp/status/2103707567520374796) What happens when your prompt is too long: - Attention dilution: The model spreads its focus across 50 rules that have nothing to do with the bug you are fixing right now. - Token tax: You pay for those 60 lines of text on every single turn of the conversation, even when you just ask a 5-word question. - Rule collisions: When two long rules slightly contradict each other, the AI guesses randomly instead of asking you. ### Flipping the Pyramid (How Top Builders Structure Agents) A few friends of mine who write some of the sharpest essays on X have a rule for writing: if you try to emphasize every sentence on a page with bold text, nothing is emphasized. Yesterday, I posted a quick thought about why developer prompts fail after turn three, and a builder named Martin (@martin_ict_algo) replied with a diagram that every developer using AI should print out and tape to their monitor: [https://x.com/poorvith_mp/status/2113502678747644037](https://x.com/poorvith_mp/status/2113502678747644037) [https://x.com/martin_ict_algo/status/2113528945799008668](https://x.com/martin_ict_algo/status/2113528945799008668) Martin pointed to the "Trust Pyramid" (originally shared by engineer Lauren Tan). Look at the order of priority from most reliable to least reliable: 1. Code structure and strict types first. 2. Linters, automated tests, and CI checks second. 3. Small, on-demand skill files third. 4. English prose rules in the prompt dead last. Almost every developer I know builds this pyramid completely upside down. We write loose, untyped code with zero automated checks, and then we try to save ourselves by writing a 500-word English essay in `CLAUDE.md` or `.cursorrules` begging the AI to write clean code. English prompts are polite suggestions; compilers and type checks are the law. If you are wasting five lines of your system prompt explaining what fields belong inside a user object, delete those five lines immediately. Put a strict TypeScript interface or Python Pydantic schema in your code instead. When the AI makes a mistake, the compiler slaps its hand in 0.2 seconds with a red error line, and the AI fixes it on the spot. ### The "How" vs. The "Never" Rule When I decided to clean up my own setup last month, I sat down and deleted 80% of my global rules file. To decide what stayed and what got deleted, I split every instruction into two buckets: - Bucket 1: The "Never" Rules (Always-On). These are the landmines that will ruin your week if the agent steps on them even once. - Never commit a live `.env` file or secret key. - Never run `git push --force` on the main branch. - Never delete a database migration file without asking me first. - Bucket 2: The "How" Rules (On-Demand Skills). These are step-by-step recipes for specific jobs you only do occasionally. - How to write a new database migration. - How to run a security audit on open ports. - How to format a release note. Your main system prompt (`AGENTS.md` or `CLAUDE.md`) should only contain Bucket 1. Keep it under 15 lines of blunt, unmistakable "Never" rules. Bucket 2 gets kicked out of the main prompt completely. Instead, put each recipe into its own small, separate file inside a local `skills/` folder. (This exact headache is why I built and open-sourced Skillary, a collection of 315 small, modular skill files that agents only open when they need them). When my agent needs to audit a database migration today, it doesn't carry database rules around in its head all morning. It opens the database skill file for thirty seconds, follows the checklist, and closes the file. Here is the Before vs. After from my own daily coding sessions after cutting 80% of my prompt: - Token usage per session: Dropped by 62% overnight because the agent stopped re-reading 50 lines of irrelevant instructions on every message. - Response speed: First-token replies feel almost instant because the base system prompt is tiny and stays cached. - Rule compliance: Because there are only a handful of hard boundaries left in the main file, the agent actually obeys them past turn ten. ### Key Takeaways: Your 3-Minute Prompt Purge Today Earlier this month, four major AI models launched within twenty days of each other, and I wrote about why I didn't switch my setup for any of them: [https://x.com/poorvith_mp/status/2106628505358533039](https://x.com/poorvith_mp/status/2106628505358533039) A brand-new frontier model won't save a cluttered workflow. If you want your current AI agent to feel twice as smart today without spending an extra dollar, open your `CLAUDE.md`, `.cursorrules`, or `AGENTS.md` file right now and do this three-step purge: 1. Delete every rule a tool can enforce. If ESLint, Prettier, Ruff, or TypeScript can catch the mistake automatically, delete the English sentence about it from your prompt. 2. Evict your "How-To" recipes into separate files. Move deployment checklists, testing guides, and styling rules into small markdown files inside a `skills/` folder so they only load when needed. 3. Cut your main file down to 10 hard "Never" laws. Short sentences. Zero fluff. Only the boundaries that protect your code and secrets from disaster. Question for you: Go open your main AI rules file (`CLAUDE.md` or `.cursorrules`) right now. How many lines long is it today, and what is the one rule in there that your AI ignores the most? Drop your line count in the replies!

Back to Blog