The systembehind my output

Headroom

Reduce what the model reads

Context economy. A graphify knowledge graph, Obsidian memory, and a grep-before-read hook let the agent load a map of the codebase instead of the whole thing. That leaves room for it to actually reason.

Ponytail

Reduce the code it writes

A lazy senior dev reflex. YAGNI first, then stdlib, then a native platform feature, then one line, then the minimum that works. Best code is the code you never write. Shortest working diff wins.

Caveman

Reduce the prose it writes

Terse output. Drop the articles, filler and hedging and you cut output tokens by about 75% with the technical content fully intact. Code and commits stay normal.

graphify

Compound what it remembers

Any input becomes a clustered knowledge graph, with HTML, JSON and an audit report. It's the backbone of a five-layer memory setup. On one project, loading that graph instead of re-feeding raw context each session dropped my Claude Code token use by roughly 71×. Knowledge sticks around instead of resetting. Method is in the writeup linked below.

Ralph Loop

Autonomous multi-session iteration

Custom. Progress lives in files and git. It rotates to fresh context near the limit and learns from failures through guardrails, with an 8-step enterprise workflow built in. Adapted from Huntley's Ralph technique.

subagent-driven-development

Parallel agents, two-stage review

Spins up a fresh subagent for each independent task, then reviews every one for spec compliance and code quality before it lands. One subagent per concern.

Strict Plan Mode

No code without a plan

Cursor rule. Every task is gated behind a real ADO work-item ID and an explicit plan before any code gets written. No drive-by edits.

AI-Native Development

Treat AI as a long-running collaborator

Meta-rule that tracks context across sessions and saves decisions. The rest of the harness hangs off it.

design-taste-frontend

Anti-slop interfaces

Global. Landing pages and portfolios that don't look templated. Audit first, real design systems, a strict pre-flight check before anything ships.

emil-design-eng

The invisible polish

Global. Emil Kowalski's take on UI polish, component design, and the animation calls that make software feel good.

Static-First Product

Static-first product model

Custom Cursor skill. Applies a static-first product model across a multi-product app, covering routing, APIs, and template-driven animation parity.

Next.js Best Practices

RSC boundaries done right

Plugin. Next.js file conventions, server and client boundaries, data and metadata patterns, image and font optimization.

Code Implementation

Secure Backend implementation

Dev rule. Pre-gate verification, mandatory security compliance, and strict layer boundaries before a line of backend code ships.

Improve Codebase Architecture

Find the deepening opportunities

Global. Scans for refactor and consolidation wins using the project's own domain language and recorded decisions, aiming at testable, AI-navigable code.

Production Refactor

Make code testable first

Subagent. Does the mechanical prod-code refactors a test strategy needs, like constructor injection and seams, before any tests get written.

Type Hygiene

No inline types, no slop

Cursor rule. Import discipline, shared-type reuse, no inline definitions, no redundant casts. Keeps the type surface clean.

Environment Config

Env mapping + auto-login

Dev rule. Subscription mapping and dynamic env resolution with auto-login across dev, preprod, prod and DR through the Azure MCP.

RCA Orchestrator

Trace by request / trace ID

Dev rule. Log-first root cause analysis for Kubernetes incidents. Resolve the environment, auto-login, then trace a request end to end.

App Insights KQL

ContainerLogV2 + AppInsights queries

Dev rule. Ready-made KQL entry points and a log schema reference for ContainerLogV2 and Application Insights.

GA4 Drop-off Analysis

Where users fall off

GA rule. Reads GA4 for funnel drop-offs, API failures, and tracking gaps, then suggests a roadmap to fix them.

test-architect → author → critic → runner

Four agents, S-grade tests

Subagent pipeline. The architect designs the strategy with no code, the author writes JUnit 5 with Mockito and AssertJ, the critic grades against an 8-dimension rubric, and the runner executes Maven and reports the coverage delta.

Test-Driven Development

Red-green-refactor, enforced

Global. Failing test first, minimum code to pass, then refactor. Features and bug fixes both.

diagnose / systematic-debugging

No flailing on hard bugs

Global and plugin. Reproduce → minimise → hypothesise → instrument → fix → regression-test, for both bugs and performance regressions.

E2E API Verification

One command, whole pipeline

Custom. Docker Compose E2E. Build images, boot services, run integration tests, surface failures. Local and CI.

Pre-PR Security Scan

Vulnerabilities before the PR

Custom. Runs an open-source dependency vulnerability scan, parses the report, and drives selective remediation by severity.

Security Checks

OWASP / MASVS checklist

Dev rule. A mandatory PASS/FAIL security checklist mapped to OWASP Top 10, ASVS, and API security. Enforced, not advisory.

Code Review Gate

AI review before commit

Dev rule. A read-only subagent review gate that runs after staging and before commit, with a bounded fix loop.

Commit Workflow

Staged, reviewed, traceable

Dev rule. Stage → review gate → conventional commit → push → security scan → PR → tracker update. Every step mandatory.

graphify

Any input → knowledge graph

Global. Builds a clustered knowledge graph from code, docs, or papers, with HTML, JSON and an audit. The backbone of low-token recall.

understand-anything

Codebase → interactive graph

Plugin, 8 skills. Map the architecture, query it in chat, analyze diffs, extract domains, and generate onboarding, all from one knowledge graph.

obsidian-workflow

Decisions into a vault

Custom. Saves architecture decisions, debug post-mortems, and ADO notes to an Obsidian vault through MCP. Memory that outlives the session.

humanizer

Strip the AI tells

Global. Strips the tells of AI prose like inflated symbolism, rule-of-three, and em-dash overuse, so the writing reads like a person wrote it. This page ran through it.