Say Goodbye to Bloated Prompts! Cursor's Dynamic Context Discovery Makes AI Coding Smarter

Dynamic Context Discovery: smarter, token‑efficient

Cursor observed that supplying too many details in the initial prompt can overload the model. Their solution is “dynamic context discovery”, where the agent requests information on demand instead of receiving a static, fully‑filled context.

Practice 1 – Treat verbose tool output as a file

When an AI runs a shell command and receives a full screen of logs, the output is written to a temporary file. The agent is informed that the result resides in the file and can preview it with tail. It reads more only if needed, preserving completeness while avoiding unnecessary token consumption.

Practice 2 – Provide a history archive during summarization

When the conversation approaches the context‑window limit, Cursor triggers a summarization step that compresses the dialogue into a concise work summary. The full conversation is also offered as a separate “history” file. If the agent detects missing details in the summary, it can retrieve the relevant portion from the archive.

Practice 3 – Adopt the open “Agent Skills” standard

Each skill package is defined as a file. The agent first learns the list of available skill files (static context) and loads the specific file only when a task requires that capability, analogous to flipping to the relevant chapter in an encyclopedia.

Practice 4 – On‑demand loading of Model Context Protocol (MCP) tools

MCP servers may expose dozens of tool descriptions, each with lengthy text. Cursor synchronizes those descriptions into a folder. The agent receives only the list of tool names as static context and fetches the detailed description from the folder when the tool is actually needed.

Result : an A/B test showed a 46.9 % reduction in total token consumption. The file‑based approach also enables the agent to detect tool‑state changes (e.g., required re‑authentication) and notify the user.

Practice 5 – Treat terminal sessions as files

All integrated terminal output is automatically synced to the local file system. The user can ask “Why did this command fail?” and the agent searches the relevant output file, eliminating the need for manual copy‑paste and simplifying inspection of long‑running logs.

File‑based interface as a simple abstraction

Cursor notes that files are not necessarily the final interface for LLM tools, but they provide a reliable, low‑complexity abstraction that avoids the risk of inventing incompatible protocols.

Original article: https://cursor.com/cn/blog/dynamic-context-discovery