RIG AI Engineering

A CLI + local web workbench that intakes, refines, and scores the prompts you hand your AI coding agents — then reads your past harness sessions to show you where your prompting is costing you.

This is the public teaser. It proves the product is real and shows you exactly what you get. The complete, production source lives in a private repo — see Get full access.

The problem

Your AI coding agent is only as good as the prompt you hand it. A vague, context-free prompt burns an expensive LLM call, returns code you have to re-prompt, and leaves no record of what actually worked. Most engineers run the same task through Claude Code or Codex three times before they get something usable — and learn nothing in the process.

The feedback loop that should exist — write prompt, run agent, learn what to do better next time — is almost never closed. There's no scoreboard for prompt quality, no record of which phrasings produced clean first-pass output, and no coach telling you that the prompt you're about to send is missing context, specificity, and a verification step.

What it does

A coach that sits in front of your agents, in your terminal and in a local browser app:

• Scores any prompt on 4 axes — Specificity, RIG Doctrine, Context, Actionability — and returns a 0–100 grade with concrete findings. rig score "fix the auth bug".
• Rewrites weak prompts with real repo context, acceptance criteria, and a verification step auto-injected. rig enhance / rig fix-prompt.
• A/B tests two prompt variants and declares a winner. rig ab-test "A" "B".
• Reads your existing harness sessions — Claude Code, Codex, OpenCode, Hermes, GSD-Pi — to surface silos, mega-sessions, and doctrine gaps in how you've actually been prompting. rig check --days 7.
• Local browser workbench — rig app --open launches a localhost app for upload-and-fix, template search, gate checks, and ProofPacket creation. 13 navigation surfaces, zero console errors (proven, below).
• 1,010 prompt templates across 10 categories, searchable. rig suggest <query>.
• Coaching diagnostics — rig coach reports your personal weaknesses and recommendations from your own session history.
• MCP server — exposes the scoring/fix/catalog tools to any MCP-capable assistant.
• Optional integrations — zsh pre-send hook (scores prompts before they reach your agent), HTTP proxy (transparently scores/rewrites), and a macOS background watcher.

Proof

Real numbers from the repository, not marketing:

• 1,010 prompt templates ship in prompt-templates/rig-prompt-templates.json — across 10 categories: rig-doctrine-execution, fleet-infrastructure, agent-orchestration, content-engineering, client-acquisition, product-code, research-intelligence, healthcare-verticals, personal-brand, diagnostic-coaching. Each is a structured object with id, category, prompt_template, variables, use_case, and a specificity_level.
• 49 test files, 1,003 test cases across the TypeScript core (src/core, src/webview). A representative run on this machine collected 595 tests with 588 passing — see the honesty note below.
• Committed Playwright proof of the local workbench at apps/rig-prompt-master/.data/playwright/navigation-audit.json: 13/13 navigation surfaces clicked and rendered, 0 console events, prompt-run output evidence = true. This is a checked-in artifact, not a screenshot.
• 50 reviewed open-source harness resources in catalogs/open-source-agent-harnesses.yaml, each with license posture, RIG role, first experiment, and load trigger.
• 10 methodology review lenses + a 100-question intake/review bank for auditing agent and harness work.

Honesty note (read it). The full test suite shows 18 files failing to load under Node 26's Vite transform in a parallel run — those same files pass clean in isolation (e.g. schemas.test.ts → 9/9). It's an environment artifact, not a product-logic failure. The product README is equally blunt about what is not yet production-proven: live credential-backed GitHub/Gitea/QNAP/Recall sync, SSO, cloud pgvector memory, worker-agents, Vercel production, and notarized desktop packaging are roadmap or read-only-local, not shipped. You're buying a working local coach, not a finished SaaS — and the repo says so in writing.

Who it's for

• IC engineers on Claude Code, Codex, OpenCode, or Hermes who want better first-pass agent output and fewer re-prompts.
• AI coaches and eng managers who need a repeatable, auditable way to score and improve prompting across a team — with a 4-axis scoreboard and trend reports.
• Anyone tired of running the same task three times because the first prompt had no context, no specificity, and no verification step.

A peek

One real prompt-template object, straight from rig-prompt-templates.json (1 of 1,010):

{
  "id": "RIG-0001",
  "category": "rig-doctrine-execution",
  "name": "IQRSQPI-{step}-a",
  "prompt_template": "Execute IQRSQPI step [STEP] for [PROJECT]. Requirements: [REQS]. Constraints: [CONSTRAINTS]. Output: [FORMAT]. Verify: [VERIF].",
  "variables": ["TASK", "REQS", "FORMAT", "VERIF", "CONTEXT"],
  "use_case": "Systematic variation a",
  "specificity_level": 4
}

And a slice of the committed workbench proof (navigation-audit.json):

{
  "appUrl": "http://127.0.0.1:8767/",
  "navPassCount": 13,
  "consoleEvents": [],
  "promptRunHasOutputEvidence": true,
  "navResults": [
    { "label": "Workbench", "exists": true, "clicked": true, "h1After": "Workbench", "changed": true },
    { "label": "Context",   "exists": true, "clicked": true, "h1After": "Context",   "changed": true }
  ]
}

The daily loop looks like this:

# Score a prompt — 0-100 across 4 axes, with findings
rig score "fix the auth bug"

# Rewrite it with repo context, acceptance criteria, a verify step
rig enhance "fix the auth bug"

# Audit how you've actually been prompting across all 5 harnesses
rig check --days 7

# Launch the local browser workbench
rig app --open

That's the surface. The scoring engine, the enhancement/context-injection logic, the session parsers for all five harnesses, the full 1,010-template library, the workbench app, the MCP server, and the V15 operator layer (catalog audit, intake packets, gates, ProofPacket templates) are all in the private repo.

Get full access

This public repo is the teaser. The complete, production source lives in a private repo.

Want it? ⭐ Star this repo, then email mike@rodgersintelligence.com or DM Mike Rodgers on LinkedIn — say which product and your use case. You'll get pricing + a private-repo invite.