A post comparing three Claude Code tools went viral this week — 116K views. gstack (Garry Tan's, 54K stars), Superpowers (Jesse Vincent's, 121K stars), and Compound Engineering (Every Inc's, 11.5K stars). The author's thesis: these aren't competitors, they're three layers. Planning, execution, review, knowledge. You need all three.
He's right about the layers. He's wrong about needing three tools.
the loc lie
Let's get this out of the way.
Garry Tan says he shipped 600K lines of production code in 60 days. 10-20K lines per day while running YC full-time. These numbers are everywhere — the viral post, his retros, every gstack pitch.
Garry builds in Rails.
We build in TypeScript/React/Next.js. Our entire monorepo — two production apps, shared packages, infrastructure — is 300K lines. Built over months, not days. Rails generates mountains of boilerplate — migrations, model files, controller scaffolding, route definitions, view templates, test fixtures. A 600K line Rails app could easily be a fraction of that in a modern TypeScript stack.
LOC is the worst metric in software engineering. Everyone knows this. The builders I've talked to who've actually tried gstack say the same thing — the structured sprint process is genuinely useful, the office-hours mode pushes back well, but the LOC numbers are "driving me nuts" and "the worst."
A codebase isn't impressive because it's big. It's impressive because it's small enough that every agent can reason about the full system.
three tools, one problem
The viral post maps four responsibilities onto three tools:
- → Decisions: gstack (/plan-ceo-review, /plan-eng-review)
- → Execution: Superpowers or CE (/ce:work)
- → Review: CE (/ce:review, 6+ Claude reviewers) + gstack (/qa)
- → Knowledge: CE (/ce:compound → docs/solutions/)
Ten steps total. Three repos. The author even warns about "process conflicts and command overlaps" when combining all three.
Our pipeline covers the same four layers with three commands:
- → /discussion → /create-plan (decisions + spec)
- → /implement (execution + review + PR prep)
- →
.context/context.md+ plan documents (knowledge)
Six months ago it was 10+ commands. Three weeks ago it was 4. Now it's 3. In six months it'll be 2 or 1. The workflow is collapsing because models keep getting better. Stacking three tools is moving in the wrong direction.
same chef, six hats
Compound Engineering's review phase sounds impressive: "dynamic reviewer ensemble, minimum 6 always-on reviewers plus conditional ones." Correctness, security, performance, testing, maintainability, adversarial. Independent reports.
They're all Claude.
Six Claude instances reviewing Claude's output with different system prompts. Same model. Same training data. Same blind spots. The viral post uses a restaurant metaphor — "a food critic, a health inspector, and a customer panel all tasting the same dish separately." But it's really the same chef wearing six hats.
Our review loop runs Codex against Claude. Different model. Different training data. Completely different opinions about code quality. Claude is too generous with its own code. Codex will straight up tell you "this is bad code" in a way Claude won't.
Cross-model adversarial review is a real second opinion. Same-model ensemble review is the cook tasting their own food six times and pretending to be six different people.
voice-first > prompt templates
The best planning advice in the viral post is a prompt: "Interview me until you have 95% confidence about what I actually want, not what I think I should want."
Clever prompt. Still a prompt.
I just talk. Record a voice note. Unstructured, conversational, thinking out loud — the same way I'd talk to a cofounder whiteboarding something. The agents handle the rambling fine, and the tangents often contain exactly the context they need that I would have forgotten to type.
When you type, you structure. When you talk, you think. The structuring is the agent's job. A prompt that asks the right questions is good. Not needing a prompt because you're just having a conversation is better.
agent-locked vs agent-agnostic
gstack, Superpowers, CE, GSD — all Claude Code skill packs. They only work if you're running Claude Code. The moment Gemini ships a better coding agent, or Codex gets a CLI, or you want Aider for a specific task, your entire workflow is gone because it was hardcoded to one agent's command system.
Pane doesn't care which agent you use. If it runs in a terminal, it runs in Pane. The slash commands live in .claude/commands, but the cockpit that runs them — session management, worktrees, diff viewer, keyboard-driven cycling between agents — works regardless of what's in each terminal.
The terminal is the universal integration layer. Everything else is vendor lock-in with a pretty README.
the compound problem is real
I'll give Compound Engineering credit: knowledge accumulation is a first-class problem most tools ignore. Agents start every session fresh. What you learned yesterday doesn't exist today. That's genuinely broken.
But their solution — spawning five subagents to extract lessons into docs/solutions/ — is heavy machinery for what should be a simple problem. Our /discussion writes decisions to .context/context.md. Every /create-plan produces a referenceable spec. The knowledge isn't in a separate "recipe binder." It's in the artifacts the workflow already produces.
The best knowledge system is the one that doesn't feel like a separate step. If you need a special /compound command to capture what you learned, your workflow isn't capturing it naturally.
spec, read, verify
I did a Q&A with Pat Hanrahan last week. Turing Award winner. Co-creator of the programmable GPU pipeline. Stanford professor for decades. I asked him about AI and programming and he compressed the whole thing into three words:
"Spec, read, verify."
Not spec, read, verify, compound, brainstorm, plan-ceo-review, plan-eng-review, office-hours, qa, ship. Three words. Three phases. The simplest version is the correct one.
The viral post identified the right layers. But the answer isn't bolting more tools together. It's collapsing into fewer commands that cover the same ground. Elegance isn't a luxury. It's how you know you found the right abstraction.
We're at three commands and heading toward one. They're at ten steps across three repos and adding more. One of these trajectories is right.
All of our Claude Code commands and slash commands are open source: github.com/Dcouple-Inc/Pane/.claude
Watch the full Pat Hanrahan Q&A: turing.rsvp
Previous posts: