What’s new
The latest guides, and what each is for.
Every guide here was published in the last few weeks. They circle one shift: you are not running a single agent any more, you are running several — and the bottleneck moved from writing code to reviewing it and knowing what your agents are actually doing.
Every guide is available in English and Polish.
Recently published
-
Software Factories: How to Build One and Keep It From Rotting
Settles the argument the industry had all year: Dex Horthy says factories fail, PostHog says they work, and both are right about different things. You get the ceiling — agent code erodes structure in 77% of trajectories — and the self-inflicted wound sitting below it, which is fixable: cap each loop with a verification oracle you trust more than the model.
Reach for it when your agents open more pull requests than the team can review, and you have to decide what it is safe to stop reading.
Read the guide -
Dr. Skill: Audit What Your Agent Actually Loads
A read-only CLI that resolves the effective skill and MCP loadout of each coding agent, then checks it against 34 issue categories: shadowed names, duplicate skills, colliding descriptions, prompt-injection surfaces, secrets sitting in MCP configs, unpinned server packages.
Reach for it when your agent reaches for the wrong tool and you cannot tell what it actually has loaded — or before you let an unfamiliar MCP server near a repo.
Read the guide -
Pruning CLAUDE.md and AGENTS.md: The Ablation Protocol
The advice inverted in July: stop adding a rule after every mistake, start deleting. Inside is what each tool really loads and when, the truncation Codex performs without a word in the TUI, a routing table for the five places a rule can live, and nine prompts that do the cutting.
Reach for it when your context file has been growing for months and the agent still ignores the rule you added back in April.
Read the guide -
Alternative IDEs and Agent Shells
A survey of everywhere you can run Claude Code, Codex and Cursor Agent — Zed and JetBrains over ACP, herdr, Conductor, Nimbalyst, Emdash, Sculptor, Warp — sorted into four shapes, with a decision table and one rule for what even qualifies: does it run your agent, or its own?
Reach for it when you are choosing where to run agents and do not want to build a workflow on a tool that is gone in six months.
Read the guide -
herdr: The Agent Multiplexer With a Scriptable API
Real PTY panes and a persistent server like tmux, but every pane classified idle, working or blocked. Underneath the TUI sits a socket API that turns out to be the more interesting half: one agent can spawn, prompt, wait on and supervise the rest of the fleet.
Reach for it when you are running six agents and one has been sitting on a permission prompt since you went for coffee.
Read the guide -
tmux for Agent Fleets: Sessions, send-keys, wait-for
Your agents create and clean up their own worktrees now, which puts every "loop over git worktree add" tutorial a layer too low. What is left is the part no agent does for you: sessions, send-keys as an agent-to-agent channel, wait-for barriers, and hooks that report exact state instead of scraping the screen.
Reach for it when you need runs to survive a closed laptop or a dropped SSH connection, and you want agents to hand work to each other.
Read the guide -
Mobbin MCP: Real Design Examples for AI Agents
Ask an agent for a paywall and you get the statistical average of every paywall in its training data. This connects it to 620,000+ screens and 130,000 user flows captured from shipped iOS and web products, so it builds from evidence instead of from the average.
Reach for it when your AI-generated interface looks like everyone else’s AI-generated interface.
Read the guide -
Taste Skill: The Anti-Slop Frontend Skill Library
A deep dive on Leonxlnx/taste-skill (~65K stars): how it constrains AI frontend generation up front — brief inference, three dials, a design-system map and a hard anti-slop ban list — plus all 13 aesthetic variants, per-agent install, and how it composes with Impeccable and the design pipeline. Verified against upstream v2.
Reach for it when every screen your agent generates arrives as the same centred card with a gradient button, and you would rather constrain the generation than correct the output.
Read the guide -
Superpowers: A Disciplined Design-to-PR Agent Workflow
A deep dive on obra/superpowers: how a 14-skill plugin turns brainstorming, worktrees, executable plans, TDD, subagent review and fresh verification into one enforced workflow — with all ten installation targets, the evidence behind it, security, and honest limits. Verified against upstream v6.1.1.
Reach for it when your agent runs are improvised from scratch each time and you want one workflow it cannot quietly skip a step in, from the first sketch to the pull request.
Read the guide -
What Claude Code Can Do That You Haven’t Tried
~40 underused Claude Code features from Gui Ferreira’s NDC AI 2026 talk, grouped into one-time setup, your daily loop, extensibility, and work beyond code. Every command was verified against the current docs rather than taken from the talk.
Reach for it when you have used Claude Code for months and suspect you are still driving it the way it worked a year ago.
Read the guide
Looking for everything else?
The documentation changelog records every new guide, lesson and content refresh, month by month, back to February 2026.