comment 0

Professional Development – 2026 – Week 46

Agentic engineering

Rethinking Claude Code Best Practices

  • Because the newer models (e.g., Opus 5.5) are so good, only put stuff in CLAUDE.md that it can’t figure out from your code. /doctor prompt-audit cleans this up.
  • Use auto mode and adjust .claude/settings.json if it’s doing things you don’t like (especially deny rules).
  • Cloud sessions give you convenience (not running on your laptop), plus it’s basically sandboxed.
  • Use hooks. Every edit runs the linter (PostToolUse). Every task completed runs tests for that task (TaskCompleted). Every session end run all tests (Stop).
  • CI is important because it runs checks on every PR. auto-fix can listen to GitHub events and watch for failures so that it can implement a fix. There’s also a “merge when ready” option.
  • At Anthropic they tend to use one orchestrator (where you tell it what needs to be done) that starts all of the sub-sessions. Fable is good for orchestrating, and Opus is good for doing the work.
  • Old best practices are still in play, but being done elsewhere. Steering is in the model now. Approving every tool call is handled by auto mode now. Saying no is done by deny rules and the environment now. Checking if it’s done is done by running tests, hooks, and CI now. Starting every session is done by an orchestrator now.

The Home Repo: The perfect multi-repo agent setup

  • TL;DR — this allows you to have a setup that looks like a monorepo from a folder perspective, but allows each sub-project to be remain as its own separate repo (because there may be different owners, pipelines, public/private access, etc.)
  • Problem: If you have a system that spans GitHub repos, to have agents work across them, you typically have skills and tooling duplicated within each repo which you’d have to keep in sync.
  • One option is to put common skills etc. in a common location on people’s machines (~/.claude). Because they’re out of sight, they may not get updated over time. Also, this location is for user-specific setups for how a given user wants to work (vs how the team wants to work). Plus, if you work on several systems, one global set of skills may not work for each system.
  • Solution: Use a home repo to share skills and settings. CLAUDE.md or AGENTS.md explains where the repos are and what they’re for. Multiple worktrees can be used when several agents are involved.
  • What the article (https://www.aihero.dev/the-home-repo-pattern) and video fail to describe well is how to set it up!
    • Make a parent folder for your project (e.g., “acme/”) and initialize a new repo here. This is the home repo.
    • Clone your N repos for that project (e.g., “acme/storefront”, “acme/payments-api”, “acme/internal-docs”), then add these folders to the new home repo’s .gitignore so you don’t double-commit files.
    • Make a home folder (e.g., “acme/home/”) to house all of the bits you want your LLM to use across other repos — AGENTS.md, CLAUDE.md, tools, shared docs, etc.

Anthropic Just Told Us How to Prompt Claude

  • Use a single prompt with a clear goal of what you want and how Claude should verify it has achieved it. Opus 5.5, medium effort.
  • Goal = spec or plan
  • If you don’t know what you’re trying to do, have Claude interview you until it can write a spec.
  • If Claude can execute something that checks in terms of pass/fail (or a number), it can keep going without you watching it.
  • For long background tasks, use /goal to specify a condition that a separate model checks after every turn.
  • Opus and Fable 5 had 80% of the system prompt removed with no measurable degradation of its coding evals.
  • Claude delegates to subagents without you needing to ask. Only ask for subagents when doing substantial, standalone work like long audits or migrations.
  • Subagents aren’t always the right move (cost, latency). A single agent with skills can outperform a subagent design on quality.
  • Use Claude cloud sessions so you can close your laptop.
  • Parallel session only pay off once a single session can finish without you. Check out Projects in Claude Code where it runs sessions for you. (Also called the Captain or Chief of Staff pattern.) A project is one long-running chat that does most of the project management for you.
  • Use code-review, which reviews your diff in a subagent. Or start a new session.
  • Every few months review your skills to see if they’re still needed.

Psychology

Why You Avoid the Things You Actually Love Doing

  • Procrastination is not reserved just for chores; you can experience procrastination of enjoyable experiences. This tends to impact highly-sensitive people or those with ADHD traits.
  • Leisure can be low-effort (scrolling, watching shows you’ve already seen) or high-effort (good book, complex film, immersive game). High-effort activities require upfront activity and active focus (which can be depleted after a tough day), and caring takes energy,
  • Occasion matching — matching an experience to a special occasion which often never comes. You place high-effort leisure on pedestal.
  • We tend to believe more free time only exists in the future. The ideal future keeps getting pushed further away.
  • Keeping a backlog/list seems natural, but we underestimate psychological reactance — rebelling against a perceived loss of autonomy and control. Putting activities on a must-finish list applies pressure that we should be doing it. It’s not play, it’s homework.
  • Hyperopia — delaying an experience because there are lower effort tasks available. Research has shown that doing something (even if under imperfect conditions) still results in a rewarding experience.
  • Stepping down from a fast, stimulating pace to a slower one is difficult. The opposite (from bored to entertained) is easier.
  • In a fast and noisy world, a completely immersive hobby helps you nourish and recharge.

Leave a Reply

Your email address will not be published. Required fields are marked *


This site uses Akismet to reduce spam. Learn how your comment data is processed.