Concept: Skill Loading Architecture

Skill Loading Architecture is the two-phase context loading mechanism employed by modern agentic LLM platforms (such as claude-code, codex, and ChatGPT Work) to manage capability tools without overwhelming model context limits.


Two-Phase Loading Mechanism

Instead of loading the complete instruction set of every installed tool or skill into context upon session initialization, agent systems execute a two-stage lazy-loading sequence:

+-------------------------------------------------------------------+
| Stage 1: Session Start (Routing Phase)                            |
| Agent loads ONLY Skill Name + Teaser Description (~20-50 tokens)  |
+-------------------------------------------------------------------+
                                  |
                        Matching Prompt Detected
                                  v
+-------------------------------------------------------------------+
| Stage 2: Task Execution (Invocation Phase)                        |
| Agent loads full skill.md, instructions, scripts & templates      |
+-------------------------------------------------------------------+

Stage 1: The Teaser Trailer (Routing Phase)

  • During initial task evaluation, the agent scans an inventory of installed skills, loading only the name and short description field from each skill’s YAML frontmatter.
  • This lightweight inventory acts like a “teaser trailer,” allowing the agent to evaluate whether a skill matches the user’s intent without consuming substantial context tokens.

Stage 2: Deep Context Hydration (Execution Phase)

  • Only when the task requirements match a skill’s description does the agent invoke the skill, dynamically reading the full skill.md file, procedural guidelines, scripts, and edge-case templates into context.

Failure Modes & Optimization Rules

Because routing decisions depend entirely on Stage 1 descriptions, poorly authored skill descriptions create severe operational failure modes:

Description FlawOperational ImpactFix / Remediation
Vague / Sparse DescriptionThe agent fails to recognize when the skill is required; the skill is never invoked (“invisible skill”).Write explicit, trigger-focused descriptions defining exact input types and task goals.
Overly Broad DescriptionThe skill description triggers on unrelated user queries, dragging massive instructions into context unnecessarily and causing token bloat.Narrow the description boundaries to specific keywords, file extensions, or workflow triggers.
Bloated FrontmatterPlacing full procedural rules inside the Stage 1 description clutters the agent’s routing layer, triggering early routing failures (especially in chatgpt-5-6 / Codex).Keep description text under 2-3 concise sentences; move procedural guidelines inside the body of skill.md.