Manifesto·May 2026·Prism

Architecture over code.
Ownership is the floor.
Maintenance is the moat.

What we are, what we’re not, and why we win.

I·The Map of the Battlefield

There is a fight happening in AI, and almost everyone in it is fighting the wrong battle.

To understand where Prism stands, we need to name the players clearly. Five categories. Hundreds of billions of dollars deployed. And not one of them does what we do.

Upstream of us

Claude · ChatGPT · Gemini

Foundation model chats

The reasoning core itself. The most powerful general-purpose AI products on Earth, and they will remain so. They are also walled gardens, your work lives inside their product, they own the experience, and when you stop paying, your context disappears.

Sideways

n8n · Zapier · Make

Workflow automation

Excellent at deterministic plumbing between services. They own the visual graph, drag a node, connect a wire, watch data flow. They do not generate architecture. They do not maintain code. They hand technical users a canvas and leave the operating burden on the user.

Sideways

Lindy · Relevance · Cassidy

Agent platforms

Templated agents inside a walled garden. You build inside their UI, your agent runs on their infrastructure, your customer data lives on their servers. The day you cancel is the day your business operation vanishes. Convenient until the moment they're not.

Sideways

Lovable · Bolt · Replit · v0

Code generators

Beautiful React components in seconds. Also the source of the “vibe code” problem the industry has been quietly enduring: spaghetti that demos beautifully in a one-shot and crumbles on day three. They optimize for the magic moment of generation, not the next nine months of operation.

Downstream of all of them

The maintenance company

Prism

The only player operating downstream of generation. We compose the foundation models, we don’t fight them. We carry the user’s operation, they don’t carry us. We sit beside their codebase forever, opening PRs, patching CVEs, evolving schemas, healing CI, for the entire life of their company.

II·The Failure Mode We Refuse to Ship

Every AI coding tool optimizes for the same moment, the user types a prompt, watches code stream onto the screen, and feels delight.

That delight is a real product experience and we will not pretend it isn’t. But here is what happens next.

The user opens the generated codebase on day three. There are seventeen ESLint warnings, a security vulnerability in a dependency the generator pinned to an old version, an auth flow that almost works, and a database schema that has no row-level security because the generator didn’t know to add it.

On day twelve, they try to add a feature, and the codebase has drifted into a shape the original LLM can no longer reason about. On day thirty, they file a support ticket. On day sixty, they cancel.

This failure mode is structural. It is not a function of which AI you use; it is a function of treating generation as the product.

If you sell the generation, you have sold a coupon to a code graveyard.

We refuse to ship this product. We do not sell generation. We sell the substrate, the ownership, and the maintenance, and generation is the entry point, not the value proposition.

III·The Three Pillars

Three load-bearing principles. Not marketing slogans, the architectural choices everything compounds against.

Architecture over code

Pillar 01

We do not generate spaghetti and call it a feature. Every Prism app is built on a production-grade scaffold that is human-engineered: row-level security on every table, encrypted credential vault, structured error handling, OAuth flows that actually work, payment integration that handles webhook races, CSP headers, rate limiting, audit logging. The AI does not write any of this. The AI only writes the domain-specific code on top: your data model, your business logic, your specific UI. The substrate is the moat. Anyone can prompt an LLM to write a Next.js app; almost no one can ship one still standing in six months. The substrate is what stands.

Ownership is the floor

Pillar 02

Every Prism app deploys to your own GitHub, your own Vercel, your own Supabase. We do not host your code. We do not own your database. We do not hold your customer data hostage. When you cancel, your app keeps running. Your domain keeps resolving. Your customers keep being served. We sever our access, we hand back the keys, we step out of the way. You owned the asset the entire time. The only thing we ever sold you was active maintenance. Lindy’s churn is catastrophic, cancel and your business operation vanishes. Our churn is graceful, you keep the app, you lose the maintenance. That asymmetry will, over years, become the most important business fact about our company.

Maintenance is the moat

Pillar 03

We are the only AI company that opens pull requests on your repository every Tuesday morning. The Atlas Janitor bumps your dependencies through a clean PR with a deterministic diff. The Security Janitor patches your CVEs daily at 5 AM UTC, never auto-merging because security is your decision. Schema Ascension lets you evolve your database in plain English and opens a PR with the migration, the snapshot, and a validator that ran BEGIN; ROLLBACK; against your live DB before asking you to merge. Auto-Heal watches your CI and pushes fix PRs within minutes of failure. None of this requires you to log into our product. We are not asking for your attention. We are asking for your repository, and paying that attention back in commits.

The substrate vocabulary lives in code, not slides. The orchestration framework underneath every Prism build is published as @prism/customization-engine on npm. See what we open and what we keep →

ChatGPT writes your code once. Prism writes it, deploys it to your GitHub, and maintains it forever.

That is the sentence. Everything else we ship is in service of making that sentence true.

IV·The Asymmetry

We operate downstream.

Foundation models are upstream of us. They generate. Workflow tools are sideways. They orchestrate. Agent platforms are sideways. They template. We are the only player operating downstream of generation.

This positioning is asymmetric in the way every great company’s positioning is asymmetric. We do not compete with Claude, ChatGPT, or Gemini for the generation moment, we compose them. We use Claude Sonnet for our LLM passes today, and if a better model lands tomorrow, we route to it without changing our value proposition. The reasoning core is a commodity input to our product, not our product itself.

This means our defensibility does not depend on training the best model, having the most compute, or raising the most capital. It depends on shipping the post-generation lifecycle nobody else is willing to ship, because shipping it requires building boring, reliable, deterministic infrastructure for years, and the venture-funded race for AGI is allergic to boring infrastructure.

We will outlast them in this lane because they will not enter it.

V·The Mechanics That Make It Real

A manifesto without product is rhetoric. Here is the catalogue.

Eight loops we already ship, each one a piece of the post-generation lifecycle no competitor has. This list will grow. Every quarter, another autonomous maintenance loop.

Phase 2.1

Dependencies

The Atlas Janitor

Weekly per-user-slot drip dispatcher. Deterministic npm registry probe. Format-preserving package.json rewriter. One PR per repo per week. Never bumps majors. Never auto-merges. Anthropic-independent, keeps running through outages.

Phase 2.2

Security

The Security Janitor

Daily CVE patcher. Three-layer major-bump guardrail. Analyst mode for advisories with no safe patch. PR-when-empty rule so a healthy repo never gets 365 spam PRs a year.

Phase 3

Schema

Schema Ascension

Type “add a published_at column to posts.” Atlas introspects the live DB, validates the migration inside BEGIN; ROLLBACK; against production, opens a PR with the migration, a snapshot, and a rationale. Twenty hard rules in the prompt. A pre-flight regex that refuses any SQL containing top-level transaction control.

Phase 11.2 + 1.1

Resilience

Auto-Heal

GitHub webhook on CI failures, Vercel webhook on deployment failures. Atomic circuit breaker capped at three attempts per error signature. Reads the logs, regenerates the offending files, pushes a fix commit. Never touches main without a PR.

Phase 13.4

Governance

The Trust Moat

Three-mode governance: Ironclad (every action gated), Earned (autonomy unlocked per tool after five consecutive approvals), Unleashed (full autonomy). Per-tool, per-provision, persisted. Rejection is a hard reset, earning trust takes weeks, losing it takes one click.

Phase 14.3

Transparency

The Glass Box

Every agent run produces a structured trace. Users pin annotations to any frame, write steering rules that inject into the system prompt on every future tick, audit what their agent did and why. Transparency is the trust play.

Phase 14.5

Authoring

The Prism Ladder

Four rungs, Describe (plain English to plan), See (the plan as an animated graph), Steer (bidirectional editor with live re-render), Live preview (sandbox runs against your example inputs). Agent authoring is a craft surface, not a wizard.

Phase 14.1

Interop

MCP-first

Every tool in our registry already speaks the Model Context Protocol. Our agents are vendor-neutral by construction. When the agent ecosystem standardizes around MCP, which is already happening, we are not retrofitting; we are already there.

VI·The Vocabulary Strategy

We will not win the agent definition war by being loudest. We win it by being the place where the answer is already written down.

Our canonical definition, the one we will publish everywhere:

An AI agent is a system where a language model decides what to do next, uses tools to act, holds state across steps, and keeps going until the goal is met or a guardrail stops it.

No mysticism. No special category of software. A program with the reasoning loop delegated to a model.

We will publish the workflow-versus-agent frame. We will publish the six primitives, Model, Tools, Memory, Planning, Reasoning loop, Environment + feedback. We will publish the spectrum from LLM call to self-improving system. We will publish the analogies, agent is a contractor, workflow is a recipe, MCP is the USB-C port, memory is the personal assistant’s notebook.

We will publish these things because Anthropic won MCP not by being the biggest company but by being the place where the spec lived. We will be the place where the agent vocabulary lives. Buyers will come to Prism because they came to our definitions first.

This is the cheapest, highest-leverage move on the entire roadmap. It requires no engineering. It requires founder writing.

VII·The Wedge Discipline

Lindy is wide. n8n is wide. ChatGPT is infinitely wide. We will not win wide.

We will win one vertical first, completely. Marketing agencies, real estate, accounting, healthcare clinics, pick one and own it for the next six months. Not “an agent builder for real estate”, “the platform every solo realtor uses to draft their post-showing follow-ups.” Templates, language, marketing, pricing, sales motion, partnerships, content, all tuned to that single buyer.

When we win that vertical, the second one is ten times easier because the playbook exists. When we win two, the third is twenty times easier. This is how vertical SaaS companies historically grow from zero to a billion in revenue: depth first, breadth second, never the reverse.

The construction industry has been formally excluded from our roadmap by founder directive. That discipline, the willingness to say no to entire industries, is what makes the depth strategy work. We will say no to construction. We will say no to a hundred other things.

Saying no is the only way to say yes loud enough to win.

VIII·The Posture

Who we are, in the smallest possible words.

not a generation company

a maintenance company.

not a walled garden

the platform that hands you the keys.

not racing the foundation models

composing them.

not selling AI

selling an Agency-in-a-Box.

not the demo

the eighteen months after the demo.

We sit downstream of every model. We compose them, we don’t fight them. We sit upstream of the user’s operation. We carry them, they don’t carry us. We sit beside their codebase forever, opening PRs, patching CVEs, evolving schemas, healing CI, surfacing weekly reports, for the entire life of their company.

Confidence compounds. Architecture compounds. Maintenance compounds. Trust, given enough Tuesdays, compounds into a moat no foundation model lab and no agent-builder platform can cross.

Architecture over code.
Ownership is the floor.
Maintenance is the moat.

That is Prism. That is the only thing we sell. And that is the company we are going to build.

Describe your idea

May 2026 · Prism · All rights reserved