Interface
Can an agent reproduce the visible product?
Modern coding agents are good at screens, forms, CRUD, charts, and familiar interaction patterns. This is usually the easiest layer, and the least defensible moat.
Methodology
Code is only one layer of a product.
We score the part an AI coding agent can build, then subtract the parts a prompt cannot manufacture: proprietary data, network effects, rights, trust, compliance, and dependable operations.
Core workflow
Does the smaller version still do the job?
We isolate the repeatable personal workflow from the vendor's full platform. A narrow tool that handles one real job is more valuable than a fake clone with dozens of dead controls.
Data access
Can the required data be obtained legally and reliably?
Search indexes, market data, maps, social graphs, licensed media, and historical datasets cannot be generated into existence. If data is the product, code is not the blocker.
Operations
Who keeps it working next month?
Retries, deliverability, queues, backups, support, abuse prevention, API migrations, and monitoring rarely appear in a demo prompt. We count them because you inherit them.
Trust & safety
Is DIY acceptable when it fails?
Security, payroll, legal, identity, finance, and regulated workflows need more than feature parity. Audits, accountability, recovery, and qualified human review change the verdict.
Verdict rules
What the labels actually mean
Strong build
A focused personal version can preserve the main value without a hard external dependency.
Scoped build
The core is achievable, but collaboration, integrations, mobile polish, or operations leave meaningful gaps.
Weak replacement
The product is primarily a network, dataset, regulated service, rights catalogue, or infrastructure operation.
The arithmetic
How the number is produced
Version 2 starts with five independently entered layer scores and a written rationale for each. The overall number is their weighted average: interface 15%, core workflow 30%, data access 20%, operations 20%, and trust and safety 15%. A reviewed community project adds 2 points; a separately reproduced implementation adds 5. The result is held between 6 and 96. Crucially, a version 2 score never reads the verdict first.
The catalogue is being migrated in public: 13 of 1093reports currently have a layer-first assessment. Every report says whether it is layer-reviewed or legacy-calibrated. Legacy records retain the earlier verdict-first formula until their five layers are reviewed; that fallback uses a 94, 76, or 49 starting point and the moat deductions below. The verdict label remains an editorial conclusion about the sensible choice, not a hidden input to a version 2 score.
Legacy moat deductions
- −22hardware
Physical devices or hardware-generated data cannot be cloned in software.
- −20compliance regulatory
Compliance, licensing, and legal accountability are core features.
- −18marketplace liquidity
The buyer and seller network cannot be generated with code.
- −18content rights
Licensed content and distribution rights are not reproducible with an LLM.
- −17proprietary data
The useful dataset is owned, accumulated, or expensive to reproduce.
- −14network effects
The value comes from the people already using it.
- −14proprietary models
Model quality and inference operations are part of the product.
- −14scale infra
Reliability at the vendor's scale is an operations problem, not a prompt.
- −10collaboration
Permissions, presence, and shared workflows are difficult to simplify.
- −9integrations
Connectors, OAuth flows, and vendor API changes require constant upkeep.
- −9brand trust
Trust, audits, and counterparties matter more than feature parity.
- −7switching costs
Years of history, configuration, and habits make migration costly.
- −4execution polish
The last 20 percent is sync, migration fidelity, speed, and edge cases.
A score is a reading of one record against these rules, not a promise about a generated demo. Two products with the same score can be unalike in every other respect; read the layer rationales and evidence before acting.
A demo is not a replacement.
Every build brief adds the parts agents tend to skip: persistent data, validation, error states, accessibility, backups, secrets, logs, tests, responsive behaviour, deployment notes, and explicit non-goals. It also bans fake integrations and hard-coded success states.
Find a build brief