seldon

Analyzes implementation plans the way Hari Seldon analyzed civilizations — by checking structural assumptions against reality before things go wrong.

Slides • Install • Usage • Focus Modes • External Judges • Output Schema

---

Feed Seldon a plan, spec, or proposal. It reads the document, inspects your codebase for evidence, and returns a verdict: approve, approve with changes, or request major revision. Every finding includes severity, evidence from your workspace, and file references with line numbers.

Works out of the box as an inline skill for Claude Code. Plug in Codex, OpenAI, or Anthropic as an external judge for true model independence. Or write your own.

Install

Claude Code plugin (recommended)

Works with Claude Code (CLI) and Claude Desktop (Cowork). No API keys or config needed for inline review.

Option A — Download from GitHub:

1. Go to the Releases page
2. Download the latest seldon-*.zip
3. Install it:

claude plugin install ./seldon-1.1.0.zip

Option B — One-liner with gh:

gh release download --repo degrammer/seldon --pattern '*.zip' && claude plugin install ./seldon-*.zip

Option C — Claude Desktop (non-technical users):

1. Download the zip from Releases
2. Open Claude Desktop
3. Go to Settings → Plugins → Install from file
4. Select the downloaded zip
5. Type /seldon my-plan.md in any conversation

After installing, the /seldon command is available in all your Claude sessions.

Other install methods

Via skills.sh:

npx skills add degrammer/seldon

Manual (clone):

# Global — available in all projects
git clone https://github.com/degrammer/seldon.git ~/.claude/skills/seldon

# Project-level — shared with your team via git
git clone https://github.com/degrammer/seldon.git .claude/skills/seldon

Usage

/seldon my-plan.md
/seldon --focus architecture docs/migration-plan.md
/seldon --focus safety spec.md supporting-context.md

What happens

1. Seldon reads your plan file (and any supporting files)
2. Inspects the workspace to verify claims — file paths, APIs, dependencies, config, schema
3. Evaluates against a rubric: repo fit, correctness, sequencing, evaluation, safety
4. Returns a structured verdict with a visual confidence bar:

🟡 Confidence  ████████████████░░░░  0.82  (moderate)

Example output

Judge: codex via judge-runner.sh
Verdict: approve_with_changes

Summary: Plan is sound but assumes a migration path that doesn't exist yet.

🟡 Confidence  ████████████████░░░░  0.82  (moderate)

Strengths:
- Clear phasing with realistic scope per step
- Good rollback strategy for the data migration

Blocking findings:

  high — Migration depends on schema v3 which hasn't been created
  Why: Step 2 cannot begin without this prerequisite
  Evidence: No v3 migration file exists in prisma/migrations/
  Refs: prisma/schema.prisma:42, docs/plan.md:18

Open questions:
- Is the external billing API rate limit sufficient for the proposed batch size?

Focus Modes

Focus modes weight the review toward specific concerns. Default is balanced.

Mode	Emphasis
balanced	All rubric dimensions evenly
architecture	Service boundaries, dependencies, migration risk, hidden integration work
evaluation	Success criteria, regression detection, testability of quality claims
product	User-visible failure modes, sequencing, scope realism
operations	Rollout, alerting, rollback, failure handling, maintenance burden
safety	Privacy, security, hallucination controls, access assumptions

/seldon --focus safety docs/auth-redesign.md

External Judges

By default, Seldon runs inline — the current agent performs the review. For true model independence, plug in an external judge.

Setup

1. Pick a judge from the judges/ directory
2. Copy it to judge-runner.sh in the skill root
3. That's it — Seldon auto-detects it

cd ~/.claude/skills/seldon  # or .claude/skills/seldon

# Option A: Codex (requires @openai/codex CLI)
cp judges/codex.sh judge-runner.sh

# Option B: OpenAI API (requires OPENAI_API_KEY)
cp judges/openai.sh judge-runner.sh

# Option C: Anthropic API (requires ANTHROPIC_API_KEY)
cp judges/anthropic.sh judge-runner.sh

To switch back to inline review, remove judge-runner.sh.

Codex