↓ scroll or arrow keys
Claude Code Plugin
Seldon — pixel art Hari Seldon hologram

Seldon

Independent plan review for AI coding agents.
Check structural assumptions against reality before things go wrong.

degrammer/seldon
01

AI writes the plan. Who checks it?

AI assistants are writing implementation plans, migration specs, architecture proposals, and project roadmaps for software-teams and builders. These plans look polished and confident — but no one is verifying they actually match reality.

Seldon is a second pair of eyes. Like having an independent reviewer read a contractor's proposal before you sign — it checks the plan against your actual project, flags what doesn't add up, and tells you what to fix before your team spends days building the wrong thing.

Less rework. Fewer surprises. Faster, safer delivery.

02

Who's this for

🎯 Founders & CTOs

Trust but verify

Your team uses AI to plan features and migrations. Seldon gives you a plain-English verdict before work starts — so you catch bad assumptions before they become expensive rework.

Read the verdict, not the code.
📋 Product Managers

Approve with confidence

When an engineer shares an AI-generated plan, Seldon tells you what's solid and what's missing. No technical background needed to understand "3 blocking findings."

Gate plans before sprint commitment.
🛠 Engineers & builders

Catch your own blind spots

You prompted an agent to write a plan. Run /seldon before you build. It checks file paths, dependencies, and sequencing against your actual codebase in seconds.

One command before you start coding.
👥 Team leads

Review gate for AI-generated work

Add Seldon to your workflow: agent writes the plan, Seldon reviews it, the team reads the verdict together. Everyone sees the risks before committing resources.

Shared decision-making, not blind trust.
03

Plans fail silently

AI agents generate confident, detailed implementation plans. But confidence is not correctness.

Phantom files

Plans reference APIs, schemas, or migrations that don't exist in the workspace.

Broken sequencing

Step 3 depends on Step 5. Prerequisites are assumed, never verified.

Blind spots

The agent that wrote the plan can't objectively evaluate its own assumptions.

04

Two paths

Without Seldon
📝 AI generates a detailed plan
👍 Team skims it, looks good, approves
🛠 Sprint starts, devs begin building
💥 Day 3: missing migration, wrong API path
🔥 Rework, scope change, blocked teammates
Deadline missed
With Seldon
📝 AI generates a detailed plan
🔍 Seldon reviews: 2 blocking findings
🔧 Fix the plan before work starts
🛠 Sprint starts on solid ground
No surprises, no rework
Shipped on time
05

How people review AI plans today

Most teams do one of these. None of them scale.

Manual review 30+ min

Senior engineer reads every line, cross-references the codebase, checks file paths by hand. Thorough but expensive.

Catches most issues — if they have time
Skim and trust 2 min

Glance at headings, check it "looks right," approve. The most common approach. Misses everything that isn't obvious.

Catches formatting issues only
Seldon 30 sec

Automated review against the real codebase. Checks every file path, dependency, and prerequisite. Returns a structured verdict.

Catches what the codebase can prove
06

Review before you build

Feed Seldon a plan. It returns a structured verdict with evidence.

approve
0.94

Clean migration plan. All referenced files exist, deps are pinned, rollback tested.

approve_with_changes
0.72

Assumes schema v3 migration exists no v3 file in prisma/migrations/

request_major_revision
0.38

3 blocking findings: missing auth middleware, circular dep in step 4, no rollback plan.

07

Understanding the verdict

approve

The plan is solid. Claims match the codebase, sequencing is correct, no missing prerequisites. Safe to start building.

approve_with_changes

Mostly sound, but has issues that should be fixed first. Non-blocking findings are suggestions; blocking findings must be addressed.

request_major_revision

Fundamental problems. Missing files, broken dependencies, impossible sequencing. The plan needs significant rework before it's actionable.

0.90 – 1.00

High confidence. Every claim was verified against the codebase. Strong evidence for the verdict.

0.70 – 0.89

Moderate. Most claims verified, but some couldn't be checked locally (external APIs, runtime behavior).

0.50 – 0.69

Low. Significant gaps in what could be verified. The plan touches areas the reviewer couldn't fully inspect.

0.00 – 0.49

Very low. The reviewer couldn't verify most claims. Treat findings as directional, not definitive.

08

LLM as Judge

A second model reads the plan, verifies claims against the workspace, and scores against a rubric — independently from the agent that authored it.

01

Ingest plan

Read plan file and any supporting docs

02

Inspect workspace

Verify file paths, APIs, deps, schema

03

Evaluate rubric

Score fit, correctness, sequencing, safety

04

Structured verdict

JSON with findings, evidence, references

09

Where Seldon fits in your workflow

Seldon doesn't replace anyone. It's a gate between planning and building.

📋

Write spec

PM / founder
🤖

AI drafts plan

Claude / agent
🔭

Seldon reviews

automated
👥

Team reads verdict

everyone
🚀

Build starts

engineers

The verdict is plain English. Everyone on the team can read it — not just engineers.

10

Focus modes

Steer the review toward what matters most. Default is balanced.

balanced

Even coverage

All rubric dimensions weighted equally. Concrete evidence over speculation.

/seldon my-plan.md
architecture

Implementation realism

Service boundaries, dependency sprawl, migration risk, hidden integration work.

/seldon --focus architecture spec.md
evaluation

Measurability

Success criteria, regression detection, testability of quality claims, observability.

/seldon --focus evaluation spec.md
product

Delivery quality

User-visible failure modes, sequencing gaps, scope realism, product risk.

/seldon --focus product spec.md
operations

Rollout durability

Ownership, alerting, rollback plans, failure handling, maintenance burden.

/seldon --focus operations spec.md
safety

Privacy & security

Hallucination controls, citation integrity, access assumptions, unsafe fallbacks.

/seldon --focus safety spec.md
11

Inline review vs. external judge

When no external judge is configured, Seldon runs inline — the same Claude session reviews the plan itself. It still checks your codebase for evidence, but the reviewer shares the same model and biases as the author.

Best practice: use a judge from a different provider. Different training data means different blind spots — that's where real bugs get caught.

Inline (default)

Zero config

Works out of the box. Good for quick checks, but shares the same blind spots as the author.

# no setup needed /seldon my-plan.md
OpenAI / Gemini

True independence

Different training = different blind spots. The recommended setup.

# activate the judge cp judges/openai.sh judge-runner.sh
Codex CLI

Deepest review

Full codebase access in a sandbox. Schema-enforced output.

# activate the judge cp judges/codex.sh judge-runner.sh
12

A Tuesday morning at a startup

📋
Sarah — Product Manager

Asks Claude to plan the new payments integration

Claude generates a 4-phase plan: new Stripe webhook, migration script, updated checkout flow, monitoring. It looks thorough. Sarah is about to approve it for the sprint.

🔍
Sarah runs Seldon

Types /seldon payments-plan.md in Claude Desktop

30 seconds later, Seldon returns request_major_revision with 2 blocking findings.

What Seldon found

Phase 2 references stripe_webhook_v2.ts which doesn't exist. Phase 3 depends on a database column (payment_intent_id) that the migration in Phase 1 never creates.

Without Seldon, the team would have discovered these gaps mid-sprint — 3 days of wasted work and a missed deadline.

Two minutes of review saved a week of rework.

13

In practice

You run
$ /seldon docs/migration-plan.md
$ /seldon --focus safety spec.md context.md
$ /seldon --focus architecture spec.md
You get
Judge: codex via judge-runner.sh
Verdict: approve_with_changes

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

Confidence ████████████████░░░░ 0.82

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

Blocking findings:
  high — Migration depends on schema v3
  Step 2 cannot begin without this prerequisite
  prisma/schema.prisma:42, docs/plan.md:18

Open questions:
• Is the billing API rate limit sufficient for batch size?
14

"At critical decision points, a holographic Seldon would appear and say: here's what you got wrong."

Named after Hari Seldon from Asimov's Foundation. He developed psychohistory — a science that predicted where civilizations would fail by checking structural assumptions against reality.

That's what /seldon does for your plans.

15

Get started

Works as a plugin for Claude Code (CLI) and Claude Cowork (desktop). No setup beyond installing the zip — no API keys, no config, no terminal skills needed.

# 1. Download the latest release from # github.com/degrammer/seldon/releases # 2. Install the zip as a plugin claude plugin install seldon.zip # 3. Review any plan /seldon my-plan.md

"If you're seeing this, here's what you got wrong."

— Hari Seldon, probably github.com/degrammer/seldon