Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

Claude Code Skills — An Agent-Harness Engineering Showcase

An engineering showcase. This repository documents eight custom Claude Code skills — small, composable programs that extend an AI coding agent with new procedures, memory, and orchestration capability. They are not prompts. They are engineering artifacts: shell daemons with re-entrancy guards, a Windows batch state machine that coordinates parallel AI agents without a single fixed sleep, a Blender script that manipulates a shader graph in code, and a 55-test pytest suite guarding a learning system's parser. What follows is a faithful, redacted account of how each was built and why.

Author: Nick Siegel · Status: active, in daily personal use · License: MIT (this showcase)


The one-paragraph version

Claude Code ships with a "skills" extension mechanism: a folder of markdown and scripts that teaches the agent a new procedure on demand. Most people who use it write a handful of one-off prompts. This repository is eight skills built as real software — with state machines, security guards, test suites, and documented production incidents — that together form a personal agent-orchestration platform: one layer decides what an AI agent should be allowed to load and remember, one layer coordinates teams of agents on work too large for one context window, and a domain layer does specific jobs (repository understanding, structured requirements-gathering, knowledge-base synthesis, even 3D art pipeline automation) inside that harness.

Across the eight skills: ~6,800 lines, spanning Bash, Python, Windows batch, and Blender's Python API, with a 55-function pytest suite, two documented production incidents with numbered fixes, and one real 8-agent, ~30–40-hour production run used as a worked example throughout.


What problem does this solve?

An AI coding agent with no structure around it fails in three predictable ways as work gets bigger and more repeated:

  1. It re-learns nothing. Every session starts cold; a correction you gave yesterday has to be given again today.
  2. It doesn't scale past one context window. A project with eight specialist roles and real dependencies between them either gets crammed into one session (thrashing) or coordinated by hand with fixed timers and hope.
  3. It loads everything, all the time. Without discipline, an agent with hundreds of available tools and skills pays a token tax on every single session just for tools it never uses that day.

These eight skills are the direct, working answer to each of those three failure modes — built, broken, fixed, and hardened through actual use rather than designed once on a whiteboard.


Why it's technically interesting

Most people's "custom AI workflows" are prompt templates. These are systems with the properties you'd expect from production software:

  • Event-driven coordination instead of timers. SE-dream-team's Windows launcher advances a multi-agent pipeline by polling a shared log for unique sentinel lines each agent must append — never a fixed sleep N. Parallel phases snapshot two sentinels and block on both. See ADR 0001.
  • A learning system hardened by real incidents, not hypotheticals. continuous-learning-v2's background observer carries fix comments tied to actual numbered issues — a re-entrancy/cooldown fix for a runaway-process memory explosion, a Windows-hang guard — plus five independent self-observation guards so the learning system can't get stuck watching itself. Backed by a 55-test pytest suite.
  • Config as a diffable, machine-written artifact. project-kickoff writes a structured "setup profile" block into every project's config specifically so a later run can diff against it instead of re-guessing — turning a one-shot interview into an idempotent, re-runnable procedure.
  • A context-budget discipline layer for multi-skill agent harnesses. brain-router classifies a task into one of nine classes before loading anything, with a hard "≤3 skills per task, never load everything" rule — the same instinct that shows up independently in this author's separate optimization of a 277-skill marketplace plugin down to a 179-skill, 8.7%-token-lighter variant (see Evidence).
  • Honest attribution, including for the skill that isn't 100% original. continuous-learning-v2 starts from an open-source base (origin: ECC, credited inspiration: the "Homunculus" community project) — the showcase says so explicitly and scopes the claimed contribution to what was actually added: a project-scoped-instinct layer plus the production hardening above. See ADR 0004.

The eight skills

Skill Layer What it does Scale
project-kickoff Platform Interviews the user, detects the stack, writes a diffable per-project Claude Code config 97 lines
continuous-learning-v2 Platform Hook-captured, project-scoped instinct learning system with an evolve-to-skill pipeline 4,313 lines / 10 files
SE-dream-team Orchestration Generates a multi-agent team package with a sentinel-paced Windows DAG launcher 869 lines / 8 files
brain-router Orchestration Task-classifying controller enforcing a load-only-what's-needed discipline across a personal agent harness 492 lines / 4 files
graphify-wrapper Domain AST + graph-clustering wrapper for cheap, structural codebase understanding 60 lines
intake Domain Bounded, structured-options elicitation that turns vague intent into an immutable artifact 215 lines
llm-wiki-maintainer Domain Karpathy-style wiki ingest/query/lint layer with multi-gate promotion thresholds 316 lines / 4 files
psx-character-pipeline Domain Six-stage Blender pipeline that scripts a PS1-style shader/material setup 443 lines / 4 files

brain-router, graphify-wrapper, intake, and llm-wiki-maintainer are four pieces of a single personal agent-harness pattern this author calls claude-brain — a controller plus three domain skills it dispatches to. They're presented individually because each is independently useful, and together in Architecture because the way they compose is the interesting part.


How it works (30-second tour)

new project ──▶ project-kickoff (interview + stack detection) ──▶ diffable per-project config
                                                                            │
large multi-role job ──▶ SE-dream-team (plan + briefs + sentinel launcher) ─┤
                                                                            │
every session ──▶ continuous-learning-v2 (hooks capture, observer learns) ─┤
                                                                            ▼
                                                        claude-brain harness (brain-router
                                                        classifies ──▶ loads only graphify-wrapper /
                                                        intake / llm-wiki-maintainer as needed)

Full detail: Architecture documentation →


The stack (and why each piece)

Concern Choice Why
Skill packaging Claude Code's native skill format (SKILL.md + scripts) The distribution unit these all had to fit — markdown-as-instructions plus real scripts underneath
Multi-agent coordination Windows batch + a shared JSONL/markdown hand-off log No message broker available in this environment; a poll-a-shared-log pattern needed zero new infrastructure
Learning capture Bash hooks (PreToolUse/PostToolUse) + a background Haiku analysis agent Hooks are deterministic (100% capture); a skill-based approach would be probabilistic (~50–80%)
Instinct storage/CLI Python Needed real data structures (dedup, clustering, promotion) and a real test suite — bash wasn't going to cut it
Codebase structure graphify (tree-sitter + Leiden clustering), wrapped Reused a purpose-built OSS tool rather than reinventing AST parsing
3D art automation Blender's Python API (bpy) The only way to script node-graph materials and Geometry Nodes reproducibly

Documentation map

Document What it covers
Engineering Story The narrative: problem → research → constraints → decisions → failures → iterations → what works
Architecture How the eight skills compose into three layers, with data-flow and state diagrams
Architecture Decision Records Six load-bearing decisions, each with options, tradeoffs, and reasoning
Evidence Real line counts, real test counts, a real production run, real token-savings numbers
Timeline How the collection grew over ~2.5 months
Interview Prep Likely questions a reviewer would ask, answered from the engineer's POV
Portfolio Summary A concise, factual summary for a resume, LinkedIn, or application
Repository Audit A self-scored hiring-manager review of this repository, with the fixes already applied
Code snippets One representative, sanitized excerpt per skill

A note on scope and honesty

This showcase follows one rule strictly: nothing here is fabricated. Every line count, test count, and incident reference is read directly from the actual skill files. Where a fact is inferred rather than confirmed by the text (why a design choice was made, an approximate date), it's marked as such. Four of these skills (brain-router, graphify-wrapper, intake, llm-wiki-maintainer) are used inside a private, unreleased AI-assistant project; this showcase presents the general-purpose harness pattern they implement and omits every file:line reference and tuned numeric threshold that was ported from that project's private internals — the engineering pattern is shown in full, the proprietary specifics are not. Personal filesystem paths have been genericized throughout.

About

Portfolio showcase: 8 custom Claude Code skills with architecture, ADRs, evidence, and interview prep

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages