Practical notes on experiments, case studies, workflows, and techniques for using Claude, Codex, and similar coding agents effectively.
How the agent skills standard turns repeated instructions into distributable task knowledge — the SKILL.md format, progressive disclosure, implicit vs. explicit invocation, scope hierarchy, testing workflow, troubleshooting, and packaging skills as plugins for team distribution.
A tour through the seven phases of the software development lifecycle where AI coding agents are changing how teams work — what to delegate, what to review, and what engineers still own.
How to run Codex agents in automated pipelines using the TypeScript SDK and Python SDK — thread lifecycle, non-interactive mode, structured output, and patterns for embedding agent runs in CI/CD.
How OpenAI's codex-action runs Codex agents in GitHub Actions with a layered security model that sandboxes the agent, drops privileges, and protects API keys from exfiltration.
How a Claude Code plugin wraps the Codex App Server to delegate reviews and rescue tasks to GPT-5.5 models — and what the forwarder subagent, background job system, and review gate patterns reveal about multi-model orchestration.
How to build lifecycle hooks for coding agents — nine events, prompt-based and command-based hooks, a JSON I/O contract, five production patterns, the Hookify rule system, and the composition model that turns agents into programmable platforms.
A hands-on guide to wiring Model Context Protocol servers into coding agents — STDIO vs HTTP transports, authentication patterns, tool filtering, timeout tuning, and six servers worth configuring today.
How Codex's plugin architecture works under the hood — marketplace catalogs, plugin manifests, installation lifecycle, permission models, and the design decisions that make an agent plugin system installable and safe.
How to think about context windows in long-running AI agent sessions — compaction strategies, AGENTS.md as persistent infrastructure, breaking complex work into compact-friendly chunks, and thread patterns that keep agents effective over hours.
LLMs don't pay equal attention across the context window — accuracy follows a U-shaped curve with a dead zone in the middle. Understanding this pattern changes how to structure prompts, place instructions, and think about 'more context.'
A practical framework for choosing sandbox modes in AI coding agents — what read-only, workspace-write, and full-write actually enforce, how sandbox inheritance works across subagents, and why per-agent overrides are defense in depth.
How Anthropic's three GitHub Actions let coding agents review PRs, triage issues, run security audits, and execute arbitrary automation inside CI/CD pipelines.
How to split code review across parallel subagents in Codex — defining TOML-based custom agents, wiring sandbox policies, and composing a review pipeline that catches what a single agent misses.
When to fan out AI agents, how to tune concurrency and depth, batch CSV processing with spawn_agents_on_csv, and the architectural patterns that make multi-agent systems predictable instead of chaotic.
How a three-level filesystem-based loading system lets agents carry dozens of capabilities while paying context cost only for what gets used.
How a manual ChatGPT cleanup loop became a Codex interceptor that narrows the gap between rough intent and executable agent work.
A self-paced quiz with 112 questions to rehearse the judgment calls the certification actually tests for.
Every MDX component available in harnez posts — props, variants, and live examples.
A hardware-led case study from Shopgro: how Lookup evolved from a grocery marketplace idea into a barcode-driven catalog capture system, pilots, grants, and hard-earned startup lessons.
A practical decision guide for deploying MCP servers covering transport selection, who pays for the LLM calls, filesystem security, and the mental model behind the three primitives.
How Claude Code subagents actually work under the hood — context isolation as the core primitive, the description field as a hidden control mechanism, and the anti-patterns that waste tokens.
The system prompt has evolved from a conversational hint into a load-bearing piece of production architecture — and it demands the same discipline we apply to config files, database schemas, and deployment pipelines.
What happens when custom tooling is built to observe how AI coding agents actually work? An experiment in understanding Claude Code through session-level observability.
When a project directory moves, Claude Code and Codex don't follow. A single script finds every stale reference across seven data stores and rewrites them to the new location.
Systematic prompt evaluation with automated test generation, dual grading systems, and measurable score progression across iterations.
How structured agent skills, iterative evaluation, and prompt refinement corrected three classes of hallucination — column invention, SQL dialect mismatches, and requirement misinterpretation — in a natural-language-to-dashboard pipeline.
Prompt caching's constraints — one-hour TTL, four breakpoints, exact-match requirements, 1,024-token minimum — are not arbitrary limits. They are a design language that pushes toward specific architectural patterns.
How a single developer built a production-ready platform for systematically evaluating AI agents through simulated conversations — from monorepo scaffold to browser-only compliance in 10 days.
When your platform has 2,000+ API endpoints across microservices and gateway layers, the surface is constantly changing. Documentation goes stale, new use cases reshape input and output schemas, and a spec-driven CLI turns API exploration into agent-ready discovery.
A reliability-first approach to choosing between predictable workflows and flexible agents, with four reusable patterns and concrete selection heuristics.
A production reference for three Claude features that meaningfully change how you build — manual cache breakpoints, auditable reasoning, and per-claim source tracking.
A tour through the execution loop that powers every agent — the five-step cycle, message types, tool execution, and the control surfaces that keep agents predictable.
How plugin-dev provides structured authoring and static validation for agent skills, while skill-creator fills the evidence gap with runtime evaluation — together covering the full lifecycle from first draft to measured improvement.
Production prompt engineering depends on iteration, edge-case design, refusal handling, and structured outputs that downstream systems can trust.
A test-driven prompt engineering system for extracting structured review data from HTML with fixture-backed validation and iterative model calibration.
Four durable properties explain where generative AI is useful, where it fails, and how to calibrate trust before using it.
Delegation, description, discernment, and diligence turn AI use from one-off prompting into repeatable collaboration.
A concise reference for system prompts, user messages, prompt structure, and the prompt components worth reaching for first.