{
  "id": 876500,
  "title": "Every Rule I Added Made It Worse: How Prompt Bloat Killed My Voice",
  "url": "https://urgent.news/2026/08/14/every-rule-i-added-made-it-worse-how-prompt-bloat-killed-my-voice",
  "topic": "world",
  "section": "World",
  "published": "2026-08-14T14:11:21.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/aws-builders/every-rule-i-added-made-it-worse-how-prompt-bloat-killed-my-voice-3ekd"
  },
  "original_language": "en",
  "account": "In a recent story, a reporter analyzed how a system prompt grew to an excessive length of 56,000 tokens due to the accumulation of rules and constraints. The original setup involved a good frontier model, a carefully crafted prompt, and iterative improvements as clichés and issues emerged. The reporter added more rules to address each failure, resulting in a complex prompt with various categories such as banned phrases, specific clichés, opener bans, register rules, anti-repetition context, work examples, meta-rules, and more. The reporter found that the prompt had become a program rather than a guide, and when rules failed to hold, they resorted to enforcing them in code with additional validators. The quality of the output deteriorated over time, becoming flatter and more cautious, lacking a distinct voice. The reporter diagnosed the issue as conflicting constraints within the prompt, causing the language model to prioritize avoiding issues over generating meaningful content. They realized that prompting could instruct a model on a specific style but couldn't make it native to the model. To address the problem, the reporter fine-tuned a model on the desired voice, reducing the prompt length by about 33 times. The key takeaway from this story is the importance of striking the right balance between prompt instructions and model capabilities, focusing on providing clear guidance rather than overwhelming the model with excessive constraints.",
  "summary": "I measured the system prompt from my old setup last week. It came to 224,833 characters, roughly 56,000 tokens, on every single call. Ninety percent of that was one block of accumulated rules about how the output should sound. I did not set out to build that. Nobody does. It grew one reasonable decision at a time, and by the end it was actively making the output worse, which I spent months…",
  "key_points": [
    "Reporter's prompt grew to 56,000 tokens due to excessive rules and constraints",
    "Prompt became a program rather than a guide, deteriorating output quality",
    "Fine-tuning model on desired voice reduced prompt length by 33 times"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}