TECHThe Smallest Agent That Works, Part 3: The Three Agents With StateTECHThe Smallest Agent That Works, Part 2: The Three Reach-Out AgentsTECHThe Smallest Agent That Works, Part 1: The Three Cheap AgentsTECHWhat MLX Got to Throw Away (That PyTorch Can't)TECHThe Unified-Memory Bet: Why On-Device Inference Stopped Being a ToyTECHEvery Useful Skill Is One of Five ShapesTECHMCP Gives You the Kitchen. Skills Are the Recipe.TECHThe Three-Level Skill, and Why Yours Is Probably One Level Too ManyTECHThe Smallest Agent That Works, Part 3: The Three Agents With StateTECHThe Smallest Agent That Works, Part 2: The Three Reach-Out AgentsTECHThe Smallest Agent That Works, Part 1: The Three Cheap AgentsTECHWhat MLX Got to Throw Away (That PyTorch Can't)TECHThe Unified-Memory Bet: Why On-Device Inference Stopped Being a ToyTECHEvery Useful Skill Is One of Five ShapesTECHMCP Gives You the Kitchen. Skills Are the Recipe.TECHThe Three-Level Skill, and Why Yours Is Probably One Level Too Many
AI × Forward
Subscribe
Tech

A working developer's toolkit for Claude Code

Ten repos that genuinely change how I work with Claude Code — and a suggested order for adopting them. Install one, use it for a week, keep only the ones that stuck. Never install a bundle.

Initial Editor·2026-04-20·4min read·887 words·26 views

Claude Code gets more useful the less you add to it.

Every extra slash command and MCP server you install becomes something you have to remember mid-flow. Most of them won't survive your next setup refactor. But a handful of repos have genuinely changed how I work across projects, across teams, across the last six months of my own tool churn.

Rule of thumb: install one, use it for a week, then decide whether to keep it. Never install a bundle.

For enforcing discipline

superpowers

A set of skills that forces structured workflow before code. The headline behavior: if there's no test for a function, the skill will refuse to write the implementation — or delete what's already there. It's annoying the first time and life-changing by week two.

Pair it with a TDD-friendly test runner and you've quietly built the guardrails your team meetings keep trying to legislate.

everything-claude-code

30+ agents, 100+ skills, dozens of slash commands. Don't install the whole thing — you'll drown. Pull three or four that match how you actually work. The kitchen-sink catalog is useful precisely because you can shop it without committing.

claude-code-best-practices

Read this one, don't install it. A set of patterns distilled from people who ship with Claude Code daily. Half an hour of reading; a couple of weeks of reversed bad habits. Worth doing before you customize anything else.

For giving Claude more context

repomix

Packs an entire codebase into a single AI-readable file with one command. The magic isn't the compression — it's that Claude stops asking clarifying questions about structure and just answers. The cost of pasting a full repo into context dropped from "annoying" to "free."

Works best on repositories under ~50k lines. Above that, prefer selective includes.

context7

Injects live, up-to-date library documentation into your prompts. If you've ever caught Claude hallucinating an API that was renamed two versions ago, this is the fix. It pulls current docs at prompt time, so you stop shipping code that references phantom functions.

claude-session-restore

Restores full context from previous sessions by walking git history. Handy when you're picking up a branch you haven't touched in a week and don't want to re-explain every decision. The generous file-size ceiling means it keeps working long after your transcript gets unwieldy.

For integrations

modelcontextprotocol/servers

The canonical MCP servers repo. The GitHub server alone is worth the install — read repos, open issues, manage PRs, all in natural language from inside Claude Code. Once you've used it, typing gh pr view manually starts to feel quaint.

claude-code-action

The official GitHub Action for running Claude Code in CI. Not demo-ware — actual PR reviews, automated, on every push. Twenty minutes to set up, and the payoff compounds every time a teammate forgets to think about edge cases.

For browsing the ecosystem

awesome-claude-code

The community's curated index. The rule for new entries is that the PR must come from Claude itself — which sets a quality bar and limits the firehose. Browse here before you build something; half the time the skill you want already exists.

awesome-claude-skills

50+ community skills grouped by category. Same spirit as awesome-claude-code, scoped to skills. Read before you build.

How to actually adopt these

The biggest mistake I see: installing six things at once. Every new tool changes how you prompt, how you think about context, and how you interpret Claude's output. When multiple variables move together, you can't tell which ones helped.

A suggested order:

  1. Start with repomix. Lowest friction, biggest immediate payoff.
  2. Add context7 next. Stops hallucinations without you having to prompt around them.
  3. Add one discipline skill — either superpowers or a carefully chosen subset of everything-claude-code. Give it two weeks before you judge it.
  4. Once those are habits, layer integrations — the MCP GitHub server, then the GitHub Action.

Four tools, deployed in order, is more useful than twelve installed the same afternoon.

Also worth a mention

The list above is about repos. A few things that aren't repos but sit right next to them — equally worth your time:

  • anthropics/anthropic-cookbook. Canonical recipes for the Anthropic API. Not Claude-Code-specific, but anyone who ends up wrapping Claude in their own tooling ends up reading this.
  • Hooks in .claude/settings.json. The single highest-leverage customization most people haven't tried. A few lines of shell in PostToolUse can auto-commit after successful edits, re-run tests, validate formatting, or kick off a deploy. Read the hooks docs before you build a custom agent — you'll often find a five-line hook replaces it.
  • A real project CLAUDE.md. The markdown file at your repo root that Claude reads on every session. Most developers treat it as a throwaway; spend an hour writing a genuine one — architectural notes, house rules, domain vocabulary — and you'll cut back-and-forth by half.
  • Cost visibility. Tools like ccusage watch token spend in real time. At daily-use cadence, a $0.50/day habit silently becomes $180/year. Putting a number on your workflow changes how you prompt.

These aren't in the main ten because they're not repos — they're adjacent practices. Equally important, different shape.

The goal

A setup that disappears into your workflow, so that when you're writing, you're writing — not managing tools.

// more in tech

see all →
Tech· 2026-05-29· 5min

The Smallest Agent That Works, Part 3: The Three Agents With State

Stateless agents fit most tasks. State is the most expensive capability you can add — it doubles your operational surface, breaks your debugging, and rewards exactly the use cases that can't survive without it. Memory, environment control, self-learning. Part 3 of three.

#agent-architecture#ai-engineering#ai-agents#system-design
Tech· 2026-05-27· 5min

The Smallest Agent That Works, Part 2: The Three Reach-Out Agents

When the cheap tiers run out, the agent has to reach beyond the model itself — into knowledge it doesn't have, tools it can't natively use, or its own previous answer. RAG, tool use, and self-critique: three patterns, three failure modes worth pricing in. Part 2 of three.

#llm#rag#agent-architecture#ai-engineering
Tech· 2026-05-26· 5min

The Smallest Agent That Works, Part 1: The Three Cheap Agents

Most agent stacks are built one tier too capable for the job. Three of the cheapest architectures — a fixed pipeline, an LLM with rule constraints, and a reasoning loop — solve more problems than the architecture diagrams admit. Part 1 of three.

#llm#agent-architecture#ai-engineering#ai-agents
Tech· 2026-05-15· 5min

What MLX Got to Throw Away (That PyTorch Can't)

Every mature framework is a museum of decisions you can't take back. MLX is interesting mostly because it started after the decisions that matter for Apple Silicon were already mistakes — and the things it threw away are the things that were quietly costing the rest of us the most.

#ai-engineering#apple-silicon#mlx#ml-frameworks
Tech· 2026-05-15· 5min

The Unified-Memory Bet: Why On-Device Inference Stopped Being a Toy

For two years the industry's default answer to every inference question has been "bigger cluster." A different hardware topology is quietly making that the wrong default for a non-trivial slice of workloads — and the framework layer that earns it is the buzzword most decks haven't caught up with yet.

#hardware#ai-infrastructure#inference#edge-ai
Tech· 2026-05-14· 5min

Every Useful Skill Is One of Five Shapes

Skills aren't a freeform format. The useful ones fit one of five shapes — sequential workflow, multi-MCP coordination, iterative refinement, context-aware selection, domain-specific intelligence. Picking the right shape is most of the design work. Picking the wrong one is most of the bugs.

#claude-code#workflow#agents#skills