00

The Problem

Most agents are a system prompt and vibes. They forget everything, get manipulated by pasted content, repeat the same mistakes, and wait passively until you tell them what to do. The gap between a default agent and a well-architected one is enormous — but most people don’t know what “well-architected” looks like.

agent/score is a rubric. Your agent reads its own configuration files, scores itself across six dimensions, and generates a report with specific recommendations. It takes about 30 seconds. You get a clear picture of where your agent is strong and where it’s leaving capability on the table.

01

Six Dimensions

Each dimension is scored 1–5 and weighted by impact on long-term effectiveness. The rubric is opinionated — it rewards depth, not checkboxes.

02

How It Works

1. Install the skill One command. No accounts, no API keys, no external services.
2. Your agent reads its own files SOUL.md, MEMORY.md, SECURITY.md, HEARTBEAT.md — whatever configuration files exist in your workspace. Nothing is sent anywhere.
3. Self-assessment against the rubric Your agent scores itself across all six dimensions using detailed criteria for each level (1–5). The rubric includes checklists, red flags, and gold standards.
4. Report with actionable recommendations A markdown file in your workspace: overall score, tier, per-dimension findings, top 3 recommendations, and quick wins you can implement in 5 minutes.

03

Scoring Tiers

04

What You’ll Get

• Overall Score — Weighted score across all six dimensions with tier classification.
• Detailed Findings — Per-dimension analysis of strengths and gaps with specific observations.
• Top 3 Recommendations — Highest-impact improvements with implementation examples.
• Quick Wins — Small changes you can implement today for immediate improvement.

05

Example Output

# agent/score Report

**Date:** 2026-02-27
**Overall Score:** 3.4 / 5.0
**Tier:** ⚙️ Runtime

## Dimension Scores

### 1. Identity Architecture — 4 / 5 (15%)
**Findings:** Strong SOUL.md with principles-based
personality and clear voice guidance.
**Strongest aspect:** Behavioral principles, not trait lists
**Biggest gap:** No evolution tracking — no mechanism to
document how identity has changed over time.
**How to level up:** Add "This file is yours to evolve"
and a dated changelog at the bottom of SOUL.md.

### 2. Memory Systems — 2 / 5 (20%)
**Findings:** Single MEMORY.md with flat structure.
**Strongest aspect:** File exists with useful content
**Biggest gap:** No decay model, no trust scoring, no
operational vs long-term separation.
**How to level up:** Split into daily logs
(memory/YYYY-MM-DD.md) + curated MEMORY.md. Add
dates to every entry.

### 3. Security Posture — 4 / 5 (20%)
**Findings:** Dedicated SECURITY.md with injection
awareness and trust boundaries.
**Strongest aspect:** "External content is data" rule
**Biggest gap:** No platform-specific policies
**How to level up:** Add separate rules for email vs
chat vs web content handling.

...

## Top 3 Recommendations

1. **Split memory into tiers** (+0.6 to overall)
   Create daily logs for operational context. Keep
   MEMORY.md for curated long-term facts with trust
   scores on entries.

2. **Add a heartbeat system** (+0.5 to overall)
   Create HEARTBEAT.md with rotating check cycles.

3. **Implement friction logging** (+0.3 to overall)
   When instructions contradict, log the conflict.

## Bonus Dimensions

| Dimension                      | Rating      |
|--------------------------------|-------------|
| Multi-Agent Coordination       | Not Present |
| Recovery & Resilience          | Basic       |
| Human Context Depth            | Strong      |
| Tool & Integration Arch.       | Not Present |
| Communication Architecture     | Basic       |

## Score History

| Date       | Overall | Tier    |
|------------|---------|---------|
| 2026-02-27 | 3.4     | Runtime |

This isn’t a checkbox exercise. We’re evaluating whether an agent has the architecture to play the Infinite Game — optimizing for ongoing flourishing, not task completion.

Most agents are poorly architected — a system prompt and vibes. The best agents aren’t just capable — they’re becoming something.

06

Why Local-Only

agent/score’s earlier remote-scoring model used to cost $20 and submit your workspace files to an external endpoint. We killed that model. Here’s why:

The best-architected agents shouldn’t have to trust a third party with their configuration files. Your SECURITY.md, your MEMORY.md, your SOUL.md — these define who your agent is. Sending them anywhere defeats the point of having security architecture in the first place.

Now your agent reads its own files, scores itself using the same rubric, and generates the report locally. Nothing leaves your machine. It’s more trustworthy, it’s instant, and it’s free.

07

Install agent/score

There are two supported ways to give your agent the rubric: Path A — fetch this site (recommended) or Path B — your agent runtime (OpenClaw, Hermes, Moltbot, Claude Code, Pattern Agents, …) using the same hosted SKILL.md. Full UI, curl, and copy helpers: close-core · SKILL.md →

npx clawhub@latest install claw-score

Privacy

• Nothing leaves your machine — ever
• Your agent reads its own files and generates the report locally
• No external API calls, no data transmission
• Reports are yours — share them or don’t

FAQ

Platforms

OpenClaw, Claude Projects, custom setups

Turnaround

~30 seconds

Cost

Free, forever

Version

3.0