Skip to content

A framework for worker-built AI apps

Prompt to Prototype is an open, vendor-neutral, evidence-based framework and field guide for a typical business that lets its workers build apps with AI, often called vibe-coded or citizen-developed apps. It shows how to keep those apps safe, secure, durable, supportable and documented across their whole lifecycle, from idea to retirement, along with the platform, operating model and governance that hold it together. It's written for leaders, IT, security, platform teams and the builders themselves. There's one recommendation for a typical business, with no company-size tiers, and every claim is sourced and dated.

Who this is for

Anyone deciding how a business should handle apps its staff build with AI. Pick your role below, or start with the 15-minute path. New to the term? Read what vibe coding means for a business. Short answers to common questions are in the FAQ.

Where should you start? Choose your role

More paths, including data and privacy owners and situations such as "we already have apps nobody tracks", are on start here: paths by role. To see the whole lifecycle worked through on realistic apps, read the field guides, which are illustrative composites covering all four risk tiers.

Start here in 15 minutes

If someone has built an app, or wants to, these three steps tell you what it needs.

  1. Answer the six tier questions. They ask about the app's data, its users, its reach beyond company sign-in, how much the business depends on it, whether it helps make decisions about people, and what actions it can take. The highest answer sets the risk tier: Sandbox, Team, Important or Critical. Read the six tier questions.
  2. Go through intake. The intake form asks the same questions, records the app in the registry and tells the builder what the tier requires. Read get a worker-built app approved.
  3. Run the go-live checklist before other people rely on the app. It covers ownership, access, data, security, testing, support and rollback, scaled to the tier. Read the go-live checklist.

Downloadable templates

Most templates are CSV files you can open in any spreadsheet tool; the pilot scope is a plain-text file. The page that explains each one is in the last column, in lifecycle order. The templates page also lists the worksheets and workshops.

Template What it's for Explained on
Discovery checklist Sources to search for the apps you already have Find existing apps
Baseline record The measures to record before a rollout, with source and confidence Baseline before you build
Pilot measures and pilot scope What a 90-day pilot measures, and its one-page scope Pilot the paved road
App record, builder core sheet The 12 fields a builder fills in to register an app App record template
App record, full registry sheet Every registry column, grouped by who fills it in App record template
App record, AI and agents sheet and audit and legal sheet The extra fields for AI features, agents and audit-scope apps App record template
App record field dictionary What every registry field means and who fills it in App record template
Decision log One row per decision record: downgrades, exceptions, approvals and program decisions Decision record template
Gate matrix What each risk tier requires at every lifecycle stage Risk tiers
Data class matrix Which data classes may go into which tools and apps Can I use company data in an AI app?
Test cases Real examples with expected answers, plus access checks Test your AI-built app
AI feature eval set Test cases and expected answers for an app that calls an AI model Evals for AI features
Verification self-check The builder's check before each release Check an AI-built app before release
Accessibility checks The checks for apps that people outside the business use Production ready
Go-live checklist The checks an app passes before people rely on it Production ready
Change checklist What to check before each change ships Change a worker-built app safely
Graduation readiness check What IT checks before it co-owns an app Graduate an app to IT
Retirement checklist Shutting an app down and handling its data and access Retire a worker-built app
Security checklist Builder self-checks and technical reviewer checks, by section Security checklist
MCP and agent baseline Minimum controls before an agent or tool server touches company data MCP and agents
Coding agent permissions baseline The eight controls for agents used to build apps Coding agents
Tool scorecard Compare candidate AI tools against the minimum standard Approve tools
Contract terms checklist Terms to require from an AI vendor before you sign Approve tools
Hosting decision table Which hosting pattern each tier may use Where to host worker-built apps
Low-code vs. AI app builders Comparing the builder types side by side Low-code vs. AI app builders
AI spend budget Caps and alerts for AI tool and model spend Control AI spend
Incident tabletop scenarios Four scenarios and their twists for a two-hour exercise Incident tabletop
Program metrics Registry health, value and cost measures, with sources and cadence Metrics for a program
ROI cost and value model Per-app and program costs against measured value Citizen developer ROI
Standards crosswalk Each standard mapped to a lifecycle stage and the page that covers it Standards crosswalk

Open data. The evidence pages publish their tables as CSV too: statistics, incidents, productivity studies, AI code security studies and success patterns.

Printable versions (DOCX and PDF). These pages also come as Word and PDF files, generated from the page itself:

Page Downloads
Risk tiers for worker-built apps DOCX, PDF
App registry template DOCX, PDF
Risk tier workshop for existing apps DOCX, PDF
Write a build spec DOCX, PDF
Check an AI-built app before release DOCX, PDF
Make a vibe-coded app production ready DOCX, PDF
Retire a worker-built app DOCX, PDF
AI acceptable use policy template DOCX, PDF
Compliance for vibe-coded apps DOCX, PDF
Standards crosswalk DOCX, PDF
Incident tabletop DOCX, PDF

What does the evidence say?

As of October 2026, four findings shape the guidance here. The statistics page gives every figure with its source, date and method.

  • Feeling faster and being faster can differ. In a 2025 randomized trial, experienced developers using AI took 19% longer on their tasks while believing they were about 20% faster.1 METR's late-2025 rerun pointed toward a speedup, but METR itself calls that data an unreliable signal.2 So the framework measures process outcomes, not self-reported time saved. See what productivity studies show.
  • AI amplifies the practices you already have. DORA's 2025 research found that AI helps teams with strong testing and review and makes weaker ones less stable.3 Guardrails and review come first. See the principles.
  • AI app builders have exposed real data, mostly through configuration. A May 2025 vulnerability record says one AI app builder's generated projects lacked row-level security, so unauthenticated users could read or write tables (the vendor disputes it).4 In February 2026 a vibe-coded social network exposed 1.5 million API tokens through a database key in its front-end code.5 See the incident log and why vibe-coded apps expose data.
  • Most staff work where some AI is in use, and many use tools nobody approved. About 41% of US workers used generative AI for work in November 2025, and about 78% of US workers were at firms using AI (employment-weighted, November 2025).6 Counted firm by firm, adoption is lower: about 25% (24.6%) of US businesses used AI in a business function in the Census survey collected 5–18 October 2026, since small firms are most firms and adopt less.7 Separately, 2026 vendor surveys found 52% to 66% of office workers using AI tools their employer hadn't approved.89 Expect worker-built apps to exist before any policy does. See find the apps you already have.

What's different here

  • Evidence is ranked by type. Controlled studies, regulator findings and post-mortems count for more than surveys, and vendor research is labelled as vendor research.
  • Failures are shown as "Don't" warnings, each tied to a documented case in the failure patterns.
  • Every page has a review date, and fast-moving claims carry an "as of" date.

Assumptions and evidence

This page adds no new assumptions; each linked page lists its own. Evidence: one randomized trial and its rerun, one large practitioner survey, government statistics, two vendor surveys, and vulnerability and vendor incident reports. Facts on this page are as of October 2026.

Last updated 22 October 2026.

  1. METR, "Measuring the impact of early-2025 AI on experienced open-source developer productivity", Jul 2025 [research], https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/, accessed 2026-10-21. Sixteen developers working on their own repositories; a preprint, not peer reviewed. ↩

  2. METR, update on its developer productivity study (late-2025 rerun), 24 Feb 2026 [research], https://metr.org/blog/2026-02-24-uplift-update/, accessed 2026-10-22. A blog report, not peer reviewed. ↩

  3. DORA, "State of AI-assisted Software Development 2025", Sep 2025 [research], https://dora.dev/research/2025/dora-report/, accessed 2026-10-21. ↩

  4. Tenable, CVE-2025-48757 record (published 30 May 2025; the vendor disputes it) [primary], https://www.tenable.com/cve/CVE-2025-48757, accessed 2026-10-21. ↩

  5. Wiz Research, "Hacking Moltbook: The AI Social Network Any Human Can Control", 2 Feb 2026 [vendor], https://www.wiz.io/blog/exposed-moltbook-database-reveals-millions-of-api-keys, accessed 2026-10-21. ↩

  6. Board of Governors of the Federal Reserve System, "Monitoring AI adoption in the U.S. economy", FEDS Notes, 3 Apr 2026 [research], https://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html, accessed 2026-10-22. The 78% is employment-weighted, from the Survey of Business Uncertainty. ↩

  7. US Census Bureau, Business Trends and Outlook Survey national data file, collection period 5–18 Oct 2026, published 22 Oct 2026 [government statistics], https://www.census.gov/hfp/btos/downloads/National.xlsx, accessed 2026-10-22. A share of firms, not weighted by employment; up from 17.7% at the start of 2026. The question wording changed in November 2025. ↩

  8. Okta, "AI Agents at Work 2026" survey (52% of 492 knowledge workers), as reported by The Register, 27 May 2026 [vendor], https://www.theregister.com/ai-ml/2026/05/27/bosses-blinded-by-confidence-about-shadow-ai-use-by-workers/5247275, accessed 2026-10-22. ↩

  9. PagerDuty, survey release on unauthorized AI tool use (66%; fieldwork by Wakefield Research), 11 Jun 2026 [vendor], https://s206.q4cdn.com/635206389/files/doc_news/PagerDuty-Report-Finds-Two-Thirds-66-of-Office-Professionals-Have-Used-Unauthorized-AI-Tools-at-Work-2026.pdf, accessed 2026-10-22. ↩