Skill System Design — Modularizing Agent Capabilities with Keyword Triggers

Mac Mini M4에서 ollama를 200% 활용하는 최적화 아키텍처

SKILL.md structure, keyword trigger tables, and pipeline composition — a practitioner's guide

Summary

  • A skill separates one agent capability into an independent module. One skill file owns one responsibility.
  • A keyword trigger table maps natural-language requests to skills automatically.
  • Connecting skills into a pipeline turns complex tasks into reproducible workflows.

Background

CLAUDE.md can tell an agent what to do. Putting how to do it into CLAUDE.md creates problems: procedures grow long, the file becomes bloated, and every session loads procedures that are irrelevant to the current task.

The skill system solves this. Each how lives in its own file and is loaded only when needed.

A mature harness carries skills for research, compilation, verification, blog publishing, Twitter drafts, deep interviews, brainstorming, code review, project diagnostics, testing, writing plans, and self-audit, among others. Each exists as an independent SKILL.md file.

ollama 프로바이더 기본 설정

Body

1. SKILL.md File Structure — Three Required Elements

Every skill file lives at .claude/skills/{skill-name}/SKILL.md. The structure is consistent.

Frontmatter: skill metadata

---
name: research
description: "Web research and information gathering for source file creation.

Trigger: 리서치, 조사, 검색, 찾아줘, research"
user-invocable: true
---

name is the unique identifier. description is what the agent reads to determine whether this skill applies. Keywords listed under Trigger are matched against user input to activate the skill. user-invocable: true allows direct invocation by the user.

Procedure: execution steps

This is the core of the skill. It specifies what to do at each step and what judgment criteria to apply. For the research skill:

1. Parse topic → confirm intent → bound scope
2. Search loop (max 3 rounds)
   - Round 1: initial search (secure 3+ independent sources)
   - Round 2+: resolve ambiguities, cross-verify
3. Classify information (fact / interpretation / estimate / opinion)
4. Counter-scenario review
5. Write search log

Explicit procedure means the agent performs the task at the same quality every time. The instruction is not "figure it out" — it is "in this order, by these criteria."

Output: result format

Specifies what format the skill's output must conform to. Without this, the agent produces different shapes of output on each run. Fixed output format is required to connect a skill to the next step in a pipeline.

2. Keyword Trigger Table — Mapping Natural Language to Skills

When a user says "look into this topic for me," which skill should the agent invoke?

A keyword trigger table in CLAUDE.md handles this mapping automatically:

| Keyword | Skill | Path |
|---|---|---|
| 리서치, 조사, 검색, 찾아줘, research | research | .claude/skills/research/SKILL.md |
| 정리, 컴파일, 소스 작성, 자료 만들어 | compile | .claude/skills/compile/SKILL.md |
| 검증, 팩트체크, 확인, 사실 확인 | verify | .claude/skills/verify/SKILL.md |
| 블로그, 게시, 발행, publish | publish | .claude/skills/publish/SKILL.md |

Design decisions worth noting:

  • Register both Korean and English keywords. The user's language choice is unpredictable.
  • Anchor keywords on verbs. "조사해줘" and "찾아줘" appear more often in natural requests than the noun "리서치" alone.
  • No overlapping keywords. If "확인" maps to two skills, the agent will be ambiguous. Each keyword resolves to exactly one skill.

Because this table lives in CLAUDE.md, the agent reads it automatically at session start. Skill activation requires only keyword matching — no separate routing logic.

3. Inter-Skill Pipelines — From Individual Skills to Workflows

Each skill is useful on its own. Connecting skills turns complex tasks into reproducible pipelines.

The core pipeline in operation:

research → compile → verify → publish

Each skill's output is the next skill's input: - Research output (raw gathered material) → Compile structures it into a source file - Compiled source file → Verify runs a four-stage fact-check - Verified source → Publish converts it to a blog post and submits it

Pipelines are defined as presets in CLAUDE.md:

| Preset | Pipeline |
|---|---|
| research | deep-interview → research → compile → verify → publish |
| quick-draft | executor → verify |
| review | code-review → verify |

The value of presets is eliminating decision cost. There is no need to reason about the processing order for each request. A complex research request follows the research preset.

4. When to Split a Skill — Separation Criteria

Too granular and maintenance overhead grows. Too coarse and skills lose their meaning. Clear criteria are required.

Signals to split:

  • Can it run independently? The verify skill can run on an existing source file with no prior research. If independent execution is meaningful, it belongs in its own skill.
  • Is it reused across pipelines? verify appears in both the research pipeline and the review pipeline. Reuse is a strong signal to separate.
  • Does the procedure exceed five steps? Longer procedures increase the probability that the agent skips intermediate steps. Splitting is safer.

Signals to merge:

  • Always executed together, and standalone execution is meaningless.
  • Both operations share identical input and output.
시행착오 / 주의사항

5. Skill Usage Tracking — memory-map.md

Skills that are built but never triggered accumulate as dead weight. A usage tracking table in memory-map.md surfaces this:

## Skill Usage Tracking
| Skill | Last Used | Frequency |
|---|---|---|
| research | 2026-04-03 | frequent |
| compile | 2026-04-03 | frequent |
| verify | 2026-04-02 | frequent |
| publish | 2026-04-01 | moderate |
| deep-interview | 2026-03-30 | occasional |
| brainstorming | — | unused |

An unused skill is one of two things: the trigger keywords are wrong, or the skill was never needed in the first place.

A rule in CLAUDE.md ensures last_used and count update automatically each time a skill triggers. Manual tracking does not persist.

6. Skill Group Responsibilities

Skills divide into three functional groups:

Group Representative Skills Role
Content Pipeline research, compile, verify Research → source file structuring → fact-check. Core workflow.
Publishing & Delivery publish, twitter-draft Blog API submission, Twitter draft delivery. Exit skills.
Design & Diagnostics deep-interview, brainstorming, writing-plans, project-doctor, and others Planning, interviewing, review, self-audit. Judgment-support skills.

Because each skill owns one responsibility, individual skill files stay short — averaging 50–90 lines.

Lessons Learned

  • Starting without skills and loading all procedures into CLAUDE.md is a common mistake. Research procedure, verification procedure, and publishing procedure all in CLAUDE.md pushed the file past 400 lines, and the agent began ignoring procedures near the bottom. After extracting to skills, CLAUDE.md dropped to the 100-line range. Only the relevant procedure loads at the relevant time.

  • Keyword collisions emerge sooner than expected. "확인" could have matched both verify and compile. Every new keyword must be checked against the existing table before registration.

  • Omitting "why" from a skill file produces mechanical compliance. A skill with procedure but no stated purpose cannot adapt when the situation calls for judgment. The Purpose section is not optional.

  • Pipelines do not need to be fully sequential. Simple requests follow the quick-draft preset and skip intermediate skills. Forcing every request through a four-stage pipeline is over-engineering.

Conclusion

The core of a skill system is separation and composition. Procedures are extracted from CLAUDE.md into independent skill files, matched to user requests via keyword triggers, and connected into pipelines that make complex workflows reproducible.

The structural advantage is extensibility. Adding a new capability requires one new skill file and one new row in the trigger table. Existing skills and CLAUDE.md require no modification.

An agent's capability is not determined by the length of CLAUDE.md. It is determined by the number and quality of its skills.

댓글

댓글 쓰기

이 블로그의 인기 게시물

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

이미지
Series overview: Series index 📚 Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. 핵심 요약 An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. 핵심 요약 CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks 📚 Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. 핵심 요약 SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… 📚 Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent 📚 Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »