The Context Front-Loading Trap
TRIGGER
General-purpose agents need domain-specific expertise to perform specialized tasks, but loading all instructions into context upfront creates bloat—most context is irrelevant to any given task, wasting tokens and potentially confusing the model with unnecessary information.
APPROACH
Anthropic's Claude Code team structured domain expertise into 'Skills'—folders containing a SKILL.md file with YAML frontmatter (name, description) plus optional auxiliary files. At startup, only skill metadata loads into the system prompt. When a task triggers a skill, Claude reads the full SKILL.md. If that file references additional context (e.g., forms.md for PDF form-filling), Claude navigates to those files only as needed. Input: user task + skill metadata in system prompt. Output: agent dynamically loads relevant instructions in layers, keeping context lean until detail is required.
PATTERN
“The front-loading trap: cramming every instruction into the system prompt degrades performance on simple tasks and bloats cost on every request. Load only metadata at startup; let the agent navigate to full instructions on demand. Context scales with task complexity, not capability breadth.”
✓ WORKS WHEN
- Agent has filesystem access to read additional context files on demand
- Domain expertise can be decomposed into hierarchical layers (overview → details → edge cases)
- Tasks vary widely in which subset of capabilities they need
- Total potential context exceeds practical context window limits (>50k tokens of instructions)
- Skill boundaries are clear enough that metadata alone signals relevance
✗ FAILS WHEN
- Agent lacks tool access to read files dynamically (API-only deployment)
- Most tasks require most of the available context anyway—progressive loading adds latency without benefit
- Domain knowledge is highly interconnected with no natural decomposition points
- Latency-sensitive applications where additional file reads are unacceptable
- Skills are too fine-grained, causing frequent context switches and navigation overhead