Fragment Selection Principle

## Overview In hakadoru.ai, **fragment management** refers to how character settings, world rules, locations, and other narrative context are included in LLM prompts. Unlike many AI writing tools that attempt automatic context injection, hakadoru.ai takes a deliberately different approach: **users explicitly select which fragments to include**. ## Why Manual Selection? The core insight is simple: **only the author knows where they are in the story**. A character setting that is critical in Chapter 1 may be irrelevant in Chapter 15. A world rule about magic systems matters during action scenes but not during a quiet conversation. The relevance of any given fragment depends entirely on the narrative timeline position — context that is impossible for an automated system to reliably determine. ### The Problems with Auto-Injection Automatic fragment injection suffers from several fundamental issues: 1. **Context window waste** — Injecting irrelevant fragments consumes tokens that could be used for actual generation 2. **Narrative confusion** — Including a character's backstory when the scene calls for mystery can break the author's intent 3. **False confidence** — Authors may assume the AI "knows" everything when it only has partial, algorithmically-selected context ### The Manual Selection Advantage By making fragment selection explicit, we achieve: - **Precision** — Only relevant context is included in each prompt - **Transparency** — Authors always know exactly what context the AI has - **Creative control** — The author remains the authority on narrative relevance - **Token efficiency** — No wasted context window on irrelevant fragments ## Implementation Fragments are stored as knowledge items and organized into folders. When composing a prompt, the user selects which items to attach. The UI makes this process fast through recent-item suggestions and folder-based browsing, but the selection itself is always a conscious author decision. ## Related Concepts - **Knowledge Assets (FR-01)** — The storage and management of fragments - **Context Adaptation (FR-07)** — How selected fragments are formatted for the LLM - **Preamble Injection** — System-level context that is always included (distinct from user fragments)