From Chat to Coding Agent: Why the Surface Changes the Rules

The cross-product captures in asgeirtj/system_prompts_leaks make comparison possible, but only after the comparison has a fair denominator. This chapter uses the prompt-stack dimensions already established, including role, authorized action, state, retrieval, output, and verification, to explain why a chat assistant, a search service, a coding agent, and a connected workspace need different rules.

A model label encourages a simple contest. Which assistant is smarter? Which prompt is more capable? Which product has the most tools? Those questions flatten different jobs into one scale. A search assistant is not a weaker coding agent because it does not edit a repository. A coding agent is not automatically better because it can run commands. Each surface is designed around a different loop between user intent, information, action, and consequence.

Compare work models, not brand names

A general chat surface is optimized for explanation, conversation, and flexible language work. Its prompt may devote substantial space to identity, tone, response structure, safety, and follow-up behavior. It can still have tools, but durable external action is not necessarily the center of its role. The product’s main failure may be an irrelevant, misleading, or poorly structured answer.

A search-oriented assistant is organized around retrieval and currentness. Perplexity’s captured prompt gives the assistant a research-oriented identity and sets expectations for thorough teaching, structured responses, knowledge-cutoff context, and current-date handling. Claude’s general prompt similarly treats search as a policy question: current products, policies, positions, and other changing facts call for retrieval, while stable knowledge may not. The product’s main failure is not only a wrong sentence. It is a confident answer built from stale or poorly selected evidence.

A coding agent operates in a more consequential environment. Claude Code, Codex, and Copilot captures describe workspaces, code-search behavior, tools, permission or authorization decisions, and validation or reporting obligations. The agent may need to read or modify files, run commands, delegate work, or interact with external services. Its main failure can be a wrong change, an unauthorized action, or a report that claims success without observing the result.

A connected workspace sits between conversation and action. It may search a drive, read a document, update a calendar, or call a service through a connector. The prompt must manage not only what the assistant can do but also which account, data source, and provider the user intends. Claude’s captured connector rules distinguish named services, directory discovery, user choice, and direct calls after an explicit selection. Here, provider choice and data scope are part of the user experience.

The comparison can be held in a compact set of dimensions:

The table does not rank the surfaces.