arete

Arete

ἀρετή — excellence earned through effort, not given.

Arete is a plugin for agentic harnesses (Claude Code, OpenCode CLI, Copilot CLI) with a clear goal: helping you build the right things, in the right way. How? By implementing a guided brainstorming framework that forces you to pause, think, and design before you implement.

Install:

# Claude Code
/plugin marketplace add jesgarram/arete
/plugin install arete@jesgarram/arete

See Cross-Platform Usage for detailed installation on other platforms.

---

Quick Start

1. Install the plugin (see above)
2. Run: /arete:brainstorm "your problem here" (Claude Code) or just ask to brainstorm a problem in (OpenCode, Copilot)
3. Follow the prompts through GROUND → EXPLORE → DECIDE → STRESS → SHIP

That's it. Arete will guide you through each phase.

---

What It Does

You: "I want to use MongoDB for this project because it's faster to set up."

Standard AI: Generates a docker-compose.yml and a schema immediately.

Arete: "The data model you described is highly relational. If we use a document store here, we'll end up managing complex relationships in application code. Are we choosing this for the data model, or for setup speed?"

Arete is your navigator — in pair-programming terms, it owns the big picture while you drive. It asks the questions your best tech lead would ask, except it's 2am and they're asleep.

---

How It Works

There are five phases. They run in order. Each has exit criteria, and you cannot skip ahead. If stress-testing reveals gaps, Arete loops back to the right phase instead of pushing forward.

When the problem is too large, Arete breaks it into focused sub-sessions that each get the full treatment.

When you need to look something up—prior art, a concept, what others have done: a researcher or teacher runs quietly in the background. You are not interrupted. The results wait in context/ until you're ready.

At the end, diagrams are drawn. Components, sequences, flows. Whatever the design requires.

---

Who This Is For

Engineers who know that code is a liability, not an asset, and who've inherited systems where nobody remembers why decisions were made. Who've sat in postmortems thinking "we knew this would happen."

If you'd rather write 100 lines that solve the problem than 1000 that look impressive, this is for you.

---

What This Is NOT

Arete won't make decisions for you. It structures the conversation you should be having and asks questions you might skip. The answers—and the judgment—are still yours.

Sometimes you'll need to pause, dig into actual specs, and come back with real numbers.

---

When to Use

Good fit	Skip it
Greenfield features	Hotfixes
Architecture decisions	Typo fixes
"Which database?" questions	"Add a button" tasks
Explaining complex topics	Anything under 30 minutes of work
Anything you'll regret in 6 months

For problems with multiple independent dimensions ("redesign auth AND migrate the DB AND change the API"), Arete can decompose them into focused sub-sessions.

---

The Workflow

flowchart LR
    Ground[GROUND] --> Explore[EXPLORE] --> Decide[DECIDE] --> Stress[STRESS] --> Ship[SHIP]

    Stress -.->|Flawed| Explore
    Stress -.->|Gaps| Decide
    Stress -.->|Reframed| Ground

Phases

Phase	Purpose	Exit Criteria
GROUND	Verify the problem exists and is worth solving	Trigger, pain, stakes, assumptions, and scope answered with specifics
EXPLORE	Generate multiple approaches to avoid tunnel vision	Multiple distinct approaches surfaced; new questions yield familiar directions
DECIDE	Select an approach and explicitly accept trade-offs	Trade-offs explicitly weighed; preference is stable
STRESS	Pre-mortem: imagine failure, then prevent it	Key failure modes probed; no new risks surfacing
SHIP	Output a verified design document	ADR + Plan saved to workspace

Each phase can loop back if gaps are found during stress-testing.

Quality Gates

GROUND has a kill switch: if stakes are vague ("it's not ideal", "nothing terrible happens"), Arete asks "The cost of inaction isn't clear. Dig deeper or park this?" This prevents wasting time on non-problems.

Throughout all phases, Arete watches for common anti-patterns:

Anti-pattern	Challenge
"It's slow"	How slow? For whom? Under what load?
"Users want X"	Which users? Did you ask them?
"Design for scale"	What's the current scale? What's the target?
"Best practice says..."	Best practice for what context?

Arete asks one question at a time and acknowledges your answer before moving on. It should feel like a conversation with a sharp colleague, not an interrogation.

---

TDD & Verification