AI Toolkit — 27 skills · 3 agents · MCP tools · coding rules — https://pirahansiah.com/ai-toolkit/
AI Engineering Stack
AI Prompt & Agent Toolkit
Twenty-seven battle-tested Claude Code skills, three code-review agents, MCP tooling, and coding rules — the exact AI engineering stack behind pirahansiah.com, organized into ready-to-use prompt packs for modern LLMs.
Prompt Packs
Each pack is a themed collection of skills and prompts. They ship as copy-paste prompt files that drop straight into Claude Code, Cursor, or any instruction-following agent.
Pack 01 — Computer Vision & Edge AI
CV Engineering
- cv-pipeline — detection, tracking, segmentation, annotation, video analysis.
- edge-deploy — Hailo, Axelera, Qualcomm, Apple Neural Engine, Jetson, Movidius.
- quantize — INT8/INT4 via NNCF, TensorRT, ONNX Runtime.
Pack 02 — Code Review & Debugging
Quality & Security
- code-reviewer — quality, security, maintainability reviews.
- debugger — errors, test failures, unexpected behavior.
- security-auditor — auth, data handling, security-sensitive code.
- diagnosing-bugs — structured loop for hard bugs and regressions.
Pack 03 — System Design
Architecture & Modeling
- codebase-design — shared vocabulary for deep modules.
- domain-modeling — pin down domain terminology and ubiquitous language.
- improve-codebase-architecture — find deepening opportunities, visual HTML report.
- grill-me / grilling / grill-with-docs — relentless interviews that stress-test a plan and produce ADRs.
Pack 04 — Agent Workflows
Autonomy & Orchestration
- autonomy-ladder — graduated autonomy from suggest-only to full-auto with audit logs.
- evaluator-optimizer — split writer from checker, loop until it passes.
- loop-builder — scheduled automations with STATE.md memory.
- handoff — compact a conversation into a pickup document.
- to-prd / to-issues / triage — conversation to PRD to grabbable issues.
Pack 05 — Engineering Discipline
Build & Refactor
- tdd — red-green-refactor, tests before code.
- modernize — update Python/C++ to latest standards, add types and tests.
- prototype — throwaway runnable prototype to validate a design.
- learn-from-failures — mine past failures into reusable rules.
Pack 06 — Writing & Communication
Human Output
- humanizer — strip signs of AI-generated writing.
- terse-output — drop ceremony, make every token count.
- teach — explain a new skill or concept clearly.
Pack 07 — Knowledge & Portfolio
Graph & Memory
- graphify — turn code, docs, papers, images, video into a knowledge graph.
- codebase-memory — query the codebase knowledge graph.
- portfolio — generate resumes, update the site, consolidate assets.
MCP Servers
MCP
codebase-memory-mcp
- Structured code queries against a persistent codebase knowledge graph — answer "how does this module work" without re-reading the tree.
Coding Rules
Rule
cpp-style
- Modern C++ conventions enforced on every change.
Rule
python-style
- Python style and linting standards.
Rule
hardware-optimization
- Performance patterns for edge and embedded targets.
Commands & Hooks
Command
/review
- One-shot code review against the project's standards.
Hooks
Session lifecycle
- code-discovery-gate, session-reminder, subagent-reminder — keep every session grounded in the codebase knowledge graph.
Built for the current generation of models. These patterns — evaluator-optimizer loops, autonomy ladders, agent handoffs, and knowledge-graph memory — are designed for the reasoning and tool-use capabilities of Claude 5, GPT-5.6, and equivalent frontier models. The full prompt packs are available on the shop page.