Features

Code generation

Archiet's pipeline converts your blueprint into a production-ready codebase in 7 steps. Each step is independently verifiable — the ZIP includes a full quality report so you know exactly what was generated and why.

The 7-step pipeline

Step 1
PRD extraction
Archiet reads your blueprint sections and Journey Wizard answers to extract a structured manifest: entity list, screen inventory, functional requirements, security constraints, and integrations. This manifest drives every downstream step.
Step 2
Capability selection
The capability selector matches your functional requirements against a catalogue of 50+ capabilities — auth, billing, workflow engine, LLM gateway, state machines, search, analytics, and more. Only matching capabilities are included in the output.
Step 3
Template rendering
Jinja2 templates for your chosen stack are rendered with your entities, capabilities, and PRD context. The output covers backend (models, routes, services, migrations), frontend (Next.js pages, components, API client), mobile (Expo/React Native), Docker, CI/CD, and a handoff pack.
Step 4
LLM stub fill
Any business logic stub that can't be pattern-matched deterministically is filled by an LLM. Deterministic fill runs first (free and instant). LLM fill only targets the remaining stubs. Security banned-patterns (eval, exec, hardcoded secrets) are stripped before writing.
Step 5
Synthetic Boot Test (SBT)
The generated app is booted in an isolated Docker container. The SBT checks that the app starts, /api/health returns 200, migrations apply cleanly, and there are no import errors. If the boot fails, generation is blocked and the error report is included in the ZIP.
Step 6
ARCHVERIFY quality scoring
Four dimensions scored on a 100-point scale: build integrity (25 pts), security posture (25 pts), production readiness (25 pts), and PRD fidelity (25 pts). A handoff-pack completeness bonus adds up to 5 extra points. Generation is blocked if the score is below 80.
Step 7
Delivery gate
The gate enforces hard blocks: must-priority functional requirements with zero generated files, empty code files (excluding __init__.py and conftest.py), and traceability coverage below the minimum. All hard blocks must clear before the ZIP is unlocked for download.

Quality score breakdown

Dimension	Points	Checks
Build integrity	25	File syntax, import validity, migration chain, package manifests
Security posture	25	httpOnly cookies, no hardcoded secrets, auth on every route, input validation
Production readiness	25	Docker, CI/CD, env config, error handling, health endpoint
PRD fidelity	25	Every entity, screen, and FR from the manifest has a corresponding file
Handoff pack bonus	+5	CLAUDE.md, ARCHITECTURE.md, DEPLOYMENT_GUIDE.md present and populated

Minimum publish score: 80

If the quality score is below 80, the delivery gate blocks the download and the quality report shows exactly which checks failed and how to resolve them. Regenerate after addressing the gaps.

Next: Stack selection Full quality scoring guide →

Loading…

The 7-step pipeline

Step 1

PRD extraction

Archiet reads your blueprint sections and Journey Wizard answers to extract a structured manifest: entity list, screen inventory, functional requirements, security constraints, and integrations. This manifest drives every downstream step.

Step 2

Capability selection

The capability selector matches your functional requirements against a catalogue of 50+ capabilities — auth, billing, workflow engine, LLM gateway, state machines, search, analytics, and more. Only matching capabilities are included in the output.

Step 3

Template rendering

Jinja2 templates for your chosen stack are rendered with your entities, capabilities, and PRD context. The output covers backend (models, routes, services, migrations), frontend (Next.js pages, components, API client), mobile (Expo/React Native), Docker, CI/CD, and a handoff pack.

Step 4

LLM stub fill

Any business logic stub that can't be pattern-matched deterministically is filled by an LLM. Deterministic fill runs first (free and instant). LLM fill only targets the remaining stubs. Security banned-patterns (eval, exec, hardcoded secrets) are stripped before writing.

Step 5

Synthetic Boot Test (SBT)

The generated app is booted in an isolated Docker container. The SBT checks that the app starts, /api/health returns 200, migrations apply cleanly, and there are no import errors. If the boot fails, generation is blocked and the error report is included in the ZIP.

Step 6

ARCHVERIFY quality scoring

Four dimensions scored on a 100-point scale: build integrity (25 pts), security posture (25 pts), production readiness (25 pts), and PRD fidelity (25 pts). A handoff-pack completeness bonus adds up to 5 extra points. Generation is blocked if the score is below 80.

Step 7

Delivery gate

The gate enforces hard blocks: must-priority functional requirements with zero generated files, empty code files (excluding __init__.py and conftest.py), and traceability coverage below the minimum. All hard blocks must clear before the ZIP is unlocked for download.

Quality score breakdown

Dimension	Points	Checks
Build integrity	25	File syntax, import validity, migration chain, package manifests
Security posture	25	httpOnly cookies, no hardcoded secrets, auth on every route, input validation
Production readiness	25	Docker, CI/CD, env config, error handling, health endpoint
PRD fidelity	25	Every entity, screen, and FR from the manifest has a corresponding file
Handoff pack bonus	+5	CLAUDE.md, ARCHITECTURE.md, DEPLOYMENT_GUIDE.md present and populated

Minimum publish score: 80

If the quality score is below 80, the delivery gate blocks the download and the quality report shows exactly which checks failed and how to resolve them. Regenerate after addressing the gaps.