◈ Open Source

my-skills

A personal Claude Code skill system — structured slash commands, specialist sub-agents, path-scoped rules, and lifecycle hooks that turn Claude into a disciplined engineering collaborator.

marketplace /plugin marketplace add danielleit241/my-skills

or clone git clone https://github.com/danielleit241/my-skills

What it provides

Slash Commands pipeline

/plan, /cook, /fix, /code-review, /learn — structured workflows that hand off artifacts between stages without re-explaining context.

Sub-agents agents

Specialist agents (scout, debugger, planner, tester, git-manager) spawned by commands for focused, bounded work with no context pollution.

Path-scoped Rules config

Design principles that load only when Claude touches a specific directory — zero noise when rules aren't relevant to the current task.

Lifecycle Hooks ops

Python scripts triggered on session start/end, pre-compact, and tool events — build checks, memory management, context injection.

Auto Memory memory

Persistent file-based memory across sessions. User preferences, feedback, project decisions, and references saved and recalled automatically.

CLAUDE.local.md personal

Gitignored personal overrides for tone, style, and shortcut preferences that stay local and never pollute the shared project config.

Core Principles

YAGNI Don't build for hypothetical future requirements. Three similar lines beat a premature abstraction.

KISS Prefer the simpler solution. No half-finished implementations, no wrapper-for-wrapper architectures.

DRY One source of truth per concern. No duplicate commands, no duplicate rules.

Brutal honesty Challenge every assumption. Diplomacy is noise; clarity is value.

Available Plugins

After adding the marketplace, install individual plugins: /plugin install <name>@my-skills

commands-bundle pipeline

/plan, /cook, /fix, /learn, /code-review, /docs-fe, /coding-level — the full slash command suite.

/plugin install commands-bundle@my-skills

agents-bundle agents

Scout, debugger, tester, planner, reviewers — sub-agents for the plan → cook → fix pipeline.

/plugin install agents-bundle@my-skills

frontend-slides frontend

Create animation-rich HTML presentations from scratch or convert PowerPoint files.

/plugin install frontend-slides@my-skills

backend-mindset backend

Clean Architecture, REST API design, TDD, event-driven patterns — language-agnostic.

/plugin install backend-mindset@my-skills

dotnet backend

C#, ASP.NET Core Minimal API, EF Core, xUnit, .NET Aspire, MassTransit.

/plugin install dotnet@my-skills

mermaidjs-v11 docs

24+ Mermaid.js v11 diagram types — flowcharts, sequence, ER, Gantt, architecture.

/plugin install mermaidjs-v11@my-skills

skill-creator meta

Create, improve, and benchmark Claude Code skills with eval pipelines.

/plugin install skill-creator@my-skills

code-review quality

Structured code review with security gates, correctness checks, and verification.

/plugin install code-review@my-skills

problem-solving reasoning

Collision-zone thinking, inversion, pattern recognition — creative unblocking techniques.

/plugin install problem-solving@my-skills

sequential-thinking reasoning

Step-by-step reasoning with revision, branching, and dynamic scope adjustment.

/plugin install sequential-thinking@my-skills

continuous-learning-v2 meta

Instinct-based learning — observes sessions, creates atomic instincts, evolves into skills.

/plugin install continuous-learning-v2@my-skills

strategic-compact productivity

Manual context compaction at logical intervals to preserve context across phases.

/plugin install strategic-compact@my-skills

◇ 01 — Structure

Repository Layout

Everything under .claude/. Each subdirectory has a single responsibility and loads only when relevant to what Claude is currently doing.

my-skills/
├── .claude/
│   ├── agents/          # Sub-agent definitions (scout, debugger, tester…)
│   ├── commands/        # Slash command implementations (/plan, /cook, /fix…)
│   ├── skills/          # Behavioral guidance loaded by the skill system
│   ├── rules/           # Path-scoped design principles (lazy-load)
│   └── hooks/           # Lifecycle event scripts (Python)
├── plans/               # Generated by /plan — one folder per feature
├── docs/                # This site (GitHub Pages)
├── CLAUDE.md            # Project-level instructions for Claude
└── CLAUDE.local.md      # Personal overrides (gitignored)

Directory Roles

.claude/agents/ Specialist sub-agents spawned by commands. Each agent has a bounded role. Never invoked directly — always called via a parent command.

.claude/commands/ Slash command Markdown files. Claude Code reads these when the user types the command. Each file defines workflow, step ordering, and which agents to spawn.

.claude/skills/ Behavioral guidance that loads when the skill system activates. Covers tone, response style, coding standards, and task-specific behavior overrides.

.claude/rules/ Path-scoped design principles. Each file activates only when Claude is working in a matched directory — agent rules for agents/, command rules for commands/.

.claude/hooks/ Python scripts triggered by lifecycle events: session start/end, tool calls, pre-compact. Used for build checks, memory management, and context injection.

plans/ Output of /plan. One subdirectory per feature, named YYMMDD-{slug}/. Contains plan.md and numbered phase files. Never edited manually during a /cook run.

◇ 02 — Commands

Slash Commands

Seven commands covering the full feature delivery lifecycle — planning, implementation, bug fixing, code review, documentation, and session configuration.

The Core Pipeline

/plan

Design

researcher · planner · reviewer

click to demo

plan.md

/cook

Implement

tester · debugger · git-manager

click to demo

patch

/fix

Repair

scout · debugger · reviewer

click to demo

Hover to rotate · Click a node to run the demo

Commands are stateless — they read files, write files, and exit. The filesystem is the only shared state between pipeline stages.

Command Reference

/plan

--fast --hard --parallel --two

Turns a natural-language description into a structured plan directory (plans/YYMMDD-{slug}/) with a plan.md and numbered phase files. Hard mode spawns researcher sub-agents before planning; Two mode produces competing plans for human review.

◎ 15K–150K · varies by mode — see breakdown below

Rplan-researcherreads codebase, returns findings

↓

Pplannerwrites plan.md + phase files

↓

Vplan-reviewerred-teams for gaps and vague steps

Token cost per mode

--fast ~15–25K planner only

--hard ~50–90K 2× researcher + planner + reviewer

--parallel ~70–120K parallel researcher pairs + planner + reviewer

--two ~90–150K 2 full competing plans

/cook

default --auto --fast --parallel

Reads an existing plan and executes it phase by phase. Spawns tester, debugger, and git-manager sub-agents per phase. Marks each phase [x] on completion. Pauses for review at each gate unless --auto is set.

◎ 25–50K · per phase · multiply by phase count

Ttesterruns success criteria, reports pass/fail

↓

Ddebuggerdiagnoses failure, applies minimal patch

↓

Ggit-managerconventional commit after each phase

Token cost per mode

default ~25–50K implement + tester + git-manager

--auto ~25–50K same pipeline, no review pauses

--fast ~20–35K skips optional review agents

--parallel ~40–80K concurrent phase implementation

/fix

default --auto --quick --review

Takes a failing test or error description. Runs scout → debugger → code-reviewer pipeline. Produces a minimal verified patch. Never touches files outside the scout's identified scope — no accidental refactoring.

◎ 15–60K · varies by mode

Sscoutlocates error, reads affected files

↓

Ddebuggerconfirms root cause, writes minimal diff

↓

Ccode-reviewerscores fix, checks for regressions

Token cost per mode

--quick ~15–30K scout + debugger, no review gate

default ~30–55K scout + debugger + code-reviewer + git-manager

--review ~35–60K full pipeline, pauses at each gate

--auto ~30–55K auto-approves ≥9.5 score, no pauses

/code-review

standalone

Standalone review without the full fix pipeline. Reads CLAUDE.md for project-specific rules then applies universal security, correctness, and quality checks. Returns scored findings with severity levels.

◎ ~10–25K · scales with diff size

/learn

memory

Saves a concept, decision, or behavioral rule to persistent memory. Writes structured Markdown files to ~/.claude/projects/ with typed frontmatter (user, feedback, project, reference) for scoped retrieval.

◎ ~3–8K · lightweight — reads session context only

/docs-fe

docs

Frontend documentation command. Generates or updates docs for UI components, pages, and frontend modules following project documentation conventions.

◎ ~15–30K · scales with number of changed endpoints

/coding-level

config

Sets the session coding level (1–3). Level 1 is beginner-friendly with explanations; Level 3 targets senior engineers — leads with trade-offs, assumes pattern fluency, and skips basics entirely.

◎ <2K · config only — loads a single level file

◇ 03 — Configuration

Configuration

Three layers: project instructions (CLAUDE.md), personal overrides (CLAUDE.local.md), and path-scoped rules that inject only when relevant.

CLAUDE.md

Project-level instructions committed to the repo. Loaded by Claude Code at every session start. Defines core principles, directory structure, and rule activation paths.

# my-skills

## Core Principles
YAGNI · KISS · DRY · Brutal honesty over diplomacy · Challenge every assumption

## Structure
.claude/
  agents/     # sub-agents spawned by commands
  commands/   # slash commands (/plan, /cook, /fix, /learn, /code-review, /docs-fe)
  skills/     # behavioral guidance loaded by skill system
  rules/      # path-scoped design principles (lazy-load)
  hooks/      # shell hooks for lifecycle events

## Rules
| File                      | Activates for           |
|---------------------------|-------------------------|
| .claude/rules/agents.md   | .claude/agents/**       |
| .claude/rules/commands.md | .claude/commands/**     |
| .claude/rules/skills.md   | .claude/skills/**       |

## Personal overrides
CLAUDE.local.md (gitignored) — personal preferences that shouldn't be committed.

CLAUDE.local.md

Gitignored personal overrides. Use this for machine-specific or workflow-specific preferences that shouldn't affect other contributors or the shared config.

Pattern: Anything in CLAUDE.local.md overrides CLAUDE.md for the same key. Keep shared config in CLAUDE.md; keep personal preferences in CLAUDE.local.md.

How Path-scoped Rules Load

Claude opens a file

The file path is matched against the rules activation table in CLAUDE.md

Matching rule file loads

Design principles for that directory are injected into the active context

Rules apply only while relevant

No cross-contamination — agent design rules never bleed into command files

settings.json

Hook registrations and tool permissions live in .claude/settings.json. Each hook entry maps an event to a command string executed by the shell.

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Bash",
      "hooks": [{ "type": "command", "command": "python .claude/hooks/build-check.py" }]
    }],
    "PostToolUse": [{
      "matcher": "*",
      "hooks": [{ "type": "command", "command": "python .claude/hooks/observe.py" }]
    }],
    "Stop": [{
      "hooks": [{ "type": "command", "command": "python .claude/hooks/session-end.py" }]
    }]
  }
}

◇ 04 — Hooks

Lifecycle Hooks

Python scripts triggered by Claude Code events. Each hook has a single responsibility and fails gracefully — a broken hook never blocks Claude from working.

Hook Inventory

build-check.py PreToolUse

Runs before Bash tool calls. Detects the project language (Node, Python, .NET, Go) and validates the build passes. Blocks on failure with a structured error report rather than letting Claude execute against a broken environment.

session-start.py Start

Fires when a Claude Code session begins. Injects project context — current branch, recent commits, open TODOs — so Claude starts every session fully oriented without the user having to re-explain the current state.

session-end.py Stop

Fires on session end. Writes a session summary to the project memory directory, capturing files modified, tools used, and key decisions as context for the next session.

pre-compact.py PreCompact

Runs before context compaction. Saves in-flight state — current task, working files, pending decisions — so nothing is lost when the context window resets mid-session.

plan-context.py Start

Detects if there is an active plan in plans/ and injects its current phase status into session context — so /cook always knows where to resume without the user pointing it at a plan file.

observe.py PostToolUse

Fires after every tool call. Logs tool usage patterns for session metrics and optionally triggers suggest_compact when context usage approaches the threshold.

suggest_compact.py PostToolUse

Monitors context window usage. When usage exceeds a configured threshold, surfaces a /compact suggestion before context pressure begins degrading response quality.

Adding a hook: Create a Python script in .claude/hooks/, then register it in .claude/settings.json under the appropriate event key. Hooks receive event data via stdin as JSON.

A skill system built for Claude Code and serious engineering

my-skills

What it provides

Core Principles

Available Plugins

Repository Layout

Directory Roles

Slash Commands

The Core Pipeline

Command Reference

Configuration

CLAUDE.md

CLAUDE.local.md

How Path-scoped Rules Load

settings.json

Lifecycle Hooks

Hook Inventory

A skill system built for
Claude Code
and serious engineering