agentic-ai-prompt-research: Study Coding Assistant Prompt Design

Overview

This repository documents reconstructed prompt patterns for agentic coding assistants, using behavioral observation and publicly shared information. It makes design ideas around prompt assembly, agent coordination, permissions, and context management easier to examine without presenting the material as an authoritative implementation or proprietary prompt dump.

It is most useful as a study reference for people building or evaluating coding agents. The documented patterns are approximations, so treat them as hypotheses about architecture rather than exact specifications.

Key Features

• Catalogs reconstructed prompts for core instructions, specialized agents, tools, and utility tasks.
• Describes coordinator and teammate patterns for multi-agent workflows.
• Examines permission explanations and staged classification for autonomous tool approval.
• Outlines modular prompt assembly, including stable and session-specific context.
• Documents memory loading and precedence patterns, including local overrides and conditional inclusion.
• Covers context-window management, session summaries, skills, and browser automation patterns.

Use Cases

• AI engineers can study architectural patterns when designing an agentic coding assistant.
• Prompt engineers can compare modular instruction, tool-description, and memory patterns while developing their own systems.
• Security researchers can use the permission and auto-approval discussions as a starting point for examining agent tool access.
• Students and educators can use the catalog to discuss orchestration and context management in coding agents.

Getting Started

No installation steps are specified. Browse the README and its linked prompt documents to explore the research.

Alternatives

• learn-claude-code: Teaches agent harness implementation through runnable Python examples, rather than researching prompt patterns and coordination mechanisms.

Considerations

• The documented prompts are reconstructed approximations based on observable behavior and public information, not verified source material or verbatim copies.
• The actual systems may differ significantly from the interpretations presented here.
• The repository describes an independent research project and disclaims affiliation with Anthropic.