Prompt ops for dev teams — runs in your repo

Treat your prompts like code.Without buying another platform.

The Prompt Evaluation & Versioning System is a drop-in framework for dev teams shipping AI features — a code repo, a dashboard template, and a Notion war-room for tracking prompt performance, running regression tests when models change, versioning prompts like code, and benchmarking cost-vs-quality across Claude, GPT, and Gemini. Runs in your repo with your API keys. No telemetry, no cloud, no vendor lock-in.

Get the Kit — $49one-time · no telemetry, no cloud · 30-day refund
Six deliverables · your repo
Eval & Versioning Repo
TS · PY
Dashboard Template
Track
Regression Test Suite
On model change
Cost-vs-Quality Benchmark
3 providers
Runs on
Claude · GPT · Gemini · your API keys · no telemetry
01.The Problem

Every team shipping AI features hits this wall by month three.

Prompt strings drift in your codebase with no version history beyond git. Eval criteria live in three different Notion docs and a stale Slack thread. The PM has opinions about what “good” looks like. Nobody has numbers.

Then a model update ships. Your accuracy quietly drops 12% on a specific customer segment. Nobody notices until support tickets start landing two weeks later.

You can build the eval pipeline yourself. Most teams do — eventually. The kit is what saves the team a focused week and three months of arguing about conventions.

The month-three wall · Symptoms
  • 01Prompts edited in PRs with no eval impact assessment
  • 02Model upgrade ships. Customers report regressions before you do.
  • 03PM keeps asking "can we switch to a cheaper model?" — no one has data
  • 04Three different prompt versions running in three different services
  • 05No rollback path when a prompt change breaks production
  • 06Eval criteria reduce to "looks good to me" reviews in Slack
02.What's Inside

Six deliverables. Engineered to ship together.

01

The Eval Framework Repo

TypeScript and Python eval harness with adapters for OpenAI, Anthropic, and Google AI SDKs. Golden dataset runner, pluggable scorers, CI integration. Drops into any existing project.

02

The Dashboard Template

Next.js dashboard for visualizing eval runs over time — accuracy trends, cost-per-1K, latency percentiles, regression alerts. Deploy to Vercel or run locally. Your data, your infra.

03

The Notion War-Room

Prompt-change RFC template, model migration decision doc, eval criteria scorecard, regression incident postmortem. Where non-engineers participate in prompt governance.

04

The Versioning Convention

Semantic versioning spec for prompts (major / minor / patch), changelog format, deprecation policy, rollback playbook. The conventions your team would otherwise argue about for three sprints.

05

The Benchmark Scaffolds

Pre-built benchmark templates for classification, extraction, summarization, generation, and tool-use tasks. Point them at your prompt and golden dataset. Get cost-vs-quality numbers across providers.

06

The Cost Models

Pricing tables for Claude (Opus / Sonnet / Haiku), GPT (4o / 4o-mini / o1), Gemini (Pro / Flash). Per-token costs surfaced in every eval run. Updated as providers change pricing.

.promptrc.ts
excerpt
export default definePromptConfig({
  prompts: {
    customer_intent_classifier: {
      version: '2.4.2',
      provider: 'anthropic',
      model: 'claude-sonnet-4-6',
      template: './prompts/customer_intent.md',
      eval: {
        dataset: './golden/intent_v3.json',
        scorer: 'llm-judge',
        thresholds: { accuracy: 0.92, hallucinationRate: 0.02 },
      },
      benchmark: ['anthropic/claude-sonnet-4-6', 'openai/gpt-4o', 'google/gemini-pro'],
    },
  },
})
03.The Four Workflows

Each one engineered for ship-day reliability. Each one repeatable.

01
~1 day first prompt

Regression Testing on Every Prompt Change

Run golden dataset evals on every prompt diff. Fail CI when accuracy, F1, or hallucination rate crosses your threshold. Catch the silent regressions before they ship to production.

02
~half day per task

Model Migration Decisions

Compare your current model against a candidate — GPT-4o vs Sonnet 4.6, Haiku 4.5 vs Flash, Opus vs o1 — on cost, quality, and latency across your real workloads. Replace vibes-based migration calls with numbers.

03
~2–3 hrs to adopt

Prompt Versioning Like Code

Semantic versioning convention (major / minor / patch), changelog structure, deprecation tracking, rollback patterns. Every prompt change has a version, a diff, an owner, and an audit trail.

04
~1 day per task family

Cost-vs-Quality Benchmarking

Pre-built scaffolds for classification, extraction, summarization, generation, and tool-use workloads. See exactly what you're trading when you move from premium to budget models — across providers.

04.A Regression Caught

Real failure mode. Model provider ships a quiet update. Accuracy craters. The kit blocks the ship.

prompt-eval · CI run · #1148
$ npx prompt-eval run --diff main..feat/intent-v242
PROMPT customer_intent_classifier
BASELINE v2.4.1 (gpt-4o-2024-08-06)
CANDIDATE v2.4.2 (gpt-4o-2024-11-20)
DATASET golden/intent_v3.json (n=247)
─────────────────────────────────────────────────────
v2.4.1v2.4.2Δ
Accuracy94.3%82.1%-12.2%🚨
F1 (weighted)0.9390.811-0.128🚨
Cost / 1K evals$1.40$1.40
p50 latency412ms398ms-3.4%
p95 latency880ms790ms-10.2%
Hallucination flags214+12🚨
─────────────────────────────────────────────────────
REGRESSION DETECTED — DO NOT SHIP
Failed categories (3): refund_request, escalation_intent, vague_query
Action: pin to v2.4.1 · Open: war-room/regression-2026-05-#1148
exit: 1

The kit ships the runner, the diff format, the threshold logic, the war-room link generation, and the CI exit codes. Drop into your pipeline. Block bad ships. Open the war-room ticket for the team to triage.

05.Provider & Stack Coverage

Claude. GPT. Gemini. Plus your stack.

First-class adapters for the three major providers. SDK compatibility with TypeScript and Python projects. CI integration with the runners your team already uses.

Provider

Anthropic

Claude Opus, Sonnet, Haiku

Native message format, tool-use eval, cost models updated per release.

Provider

OpenAI

GPT-4o, 4o-mini, o1, o1-mini

Chat completions, function calling, structured outputs, reasoning model latency tracking.

Provider

Google

Gemini Pro, Flash

Generative AI SDK, system instructions, function declarations, safety setting comparison.

Languages
TypeScript · Python
Runtimes
Node 20+ · Python 3.10+
CI
GitHub Actions · GitLab · Circle
Notion
Imports to any workspace
06.Kit, Not Platform

Your eval pipeline shouldn’t ship with vendor lock-in.

Every other tool in this category is a platform — you send them your prompts, your eval data, sometimes your customer inputs. You commit to their stack. You pay per seat or per eval. You hope they don’t pivot pricing or sunset features your workflow depends on.

This kit is the opposite trade-off on purpose. The patterns, the conventions, the runner, the dashboard, the war-room — all of it lives in your repo. The kit is a one-time $49 you fork, extend, and own.

The trade: you set up the pipeline yourself instead of clicking a button on someone’s dashboard. The reward: sovereignty.

Your repo, your infra

Eval framework runs where your code already runs. No new service, no third-party cloud, no extra deployment.

Your API keys, your bills

Provider calls go directly through your OpenAI, Anthropic, and Google accounts. No reseller markup. No surprise per-eval charges.

Your conventions, your team

Fork the kit. Extend it. Build your own scorers. The kit is patterns to own, not workflows to comply with.

07.Who This Is For

Be honest with yourself before you click buy.

Buy if
  • You ship AI features in production and prompts are load-bearing.
  • You have 10+ prompt strings drifting in your codebase.
  • You’ve already had at least one silent regression incident.
  • You’re a senior engineer, eng manager, or technical founder.
  • You want patterns to own, not a platform to depend on.
Skip if
  • You have one prompt in your codebase. You don’t need this yet.
  • You want a fully managed platform with a UI for everything.
  • You don’t have anyone on the team who reads TypeScript or Python.
  • You’re not shipping AI features to real users yet.
  • You want vendor support — this is a kit, not a service.
08.Pairs Well With

Eval the prompts, then guard the whole pipeline.

Evaluation is one layer; the rest of the reliability stack covers the others. The Prompt Regression Lab catches when a prompt change quietly breaks, the Agent Reliability Harness tests the agent that runs the prompts, and the RAG Retrieval Grader grades the retrieval feeding them.

09.Common Questions

The questions engineers actually ask before pulling out the company card.

No, and that's deliberate. Those are platforms — you commit to their stack and send them your eval data. This kit ships the patterns, conventions, and templates to build prompt evaluation inside your own repo. If you already use a platform, the kit still gives you better eval criteria, regression patterns, and Notion ops templates. If you don't, the kit lets you skip the platform decision until you actually need to scale.

Instant access · Lifetime updates · 30-day refund

Stop shipping prompts on vibes.
Ship them on numbers.

$49 once. Drop the kit into your repo. Get regression tests, model migration data, prompt versioning, and Claude/GPT/Gemini benchmarks by end of week.

Sold by RedHub AI LLC · Secured by Stripe · redhub.ai