Adaptive Skill Stack

What Problem It Solves

When an AI agent handles cross-domain tasks, it often lacks a stable way to decide which capabilities to invoke, how to combine them, and how to persist new lessons. Adaptive Skill Stack addresses this by turning request parsing, capability matching, new-capability construction, and file-backed accumulation into a traceable workflow.

How It Works

The core idea is stack first, build only when needed. Each task starts with semantic scanning and scenario positioning: domain, complexity, expected output, and dependencies. It then reads references/capability-registry.md and references/protocols.md.
- Stack mode: the request intersects with existing capabilities, so the skill activates matched modules and combines them according to the protocol.
- Build mode: if no existing capability matches, it solves the new problem using the model's native ability and distills a reusable method.
- Capability persistence: after the task, it appends capability entries, protocol entries, and optionally writes to references/knowledge/, scripts/, or assets/templates/.

Boundaries and Caveats

It fits agent scenarios that need ongoing accumulation of methods, templates, and domain knowledge, not simple one-off Q&A. The file contract is strict: SKILL.md is immutable, and structurally locked files must preserve their format. If reference files are missing, format drift occurs, or permission rules are not enforced, capability tracking and automatic loading may degrade.