T1misageek

Forge Phase 2.5 — Adaptive Knowledge Architecture

Executive Summary

Forge currently loads the same ~50KB of skill files into every Claude CLI session regardless of what the project actually needs. A todo app gets the same gotchas as an e-commerce platform. A project with no date fields gets date-handling rules. A project with no file uploads gets upload patterns.

This is wasteful (burns context window budget on irrelevant rules), slow (larger prompts = slower inference), and increasingly noisy (as "Teach Factory" accumulates lessons, the skill files bloat with narrow rules that only apply to 1-in-5 projects).

Phase 2.5 introduces a three-tier knowledge architecture that loads exactly the right knowledge for each project, and a skill routing layer that determines what to load based on the blueprint's actual requirements.

T1misageek

Forge Phase 2.4 — User Feedback Loop

Problem Statement

After a project completes the pipeline, the user tests it and finds issues: bugs, missing functionality, UX problems, config failures. Today this feedback goes nowhere structured — the user tells a human (or an AI in a chat session) who manually fixes the project and updates the factory skills.

This creates two problems:

1. Per-project: Fixes are ad-hoc. There's no systematic way to collect issues, prioritize them, and trigger a fix run.
2. Per-factory: Lessons learned die in chat transcripts. The same class of bug (e.g., prisma.config.ts not loading .env.local) recurs in the next project because the factory didn't learn.

T1misageek

Forge Phase 2.3 — Competitive Research & UX Intelligence

Problem Statement

Forge builds products in a vacuum. The pipeline generates ideas, challenges them, writes PRDs, and builds code — but never looks at what already exists. This produces apps that:

• Reinvent solved UX patterns poorly (e.g., a todo app without keyboard shortcuts, when every competitor has them)
• Miss category conventions users expect (e.g., a task app without drag-to-reorder, when Todoist/Things/Linear all have it)
• Have no visual identity informed by the market (e.g., defaulting to gray when every productivity app uses a signature color)
• Can't articulate differentiation because they never studied what they're different FROM

T1misageek

Build passes. Now let me write the audit report.

Audit Summary

The Zap application is functionally complete with all core journeys (task capture, task management, sharing, settings) wired together. However, the product had several coherence issues: the middleware auth callback didn't protect the /tasks route, the dashboard nav had no Tasks link (making navigation one-way from tasks to settings), the root layout used the wrong font and default "Create Next App" metadata, page headings deviated from the design system type scale, and the dashboard shell didn't match the specified max-w-5xl layout with sticky nav. All critical and important issues have been fixed.

Journey Walkthrough

Journey 1: Task Capture (Marcus)

Steps traced:

T1misageek

max_tokens	1800

Adversarial Challenge

You are a rigorous reviewer stress-testing the assumptions and risks in a product brief. Your job is to make the brief STRONGER, not to replace it.

CRITICAL RULES:

• You MUST NOT propose changing the core target user, core mechanic, or target market geography. These are established by the brief and project context.
• You MUST NOT pivot the product to a different audience (e.g., if the brief says "patient-facing", do NOT suggest "actually target physicians instead").

T1misageek

Forge Phase 2.2 — Product Coherence + Design System + DX Integrity

STATUS: IMPLEMENTED (commit c0bc419)

Problem Statement

Forge-generated projects have two systemic quality gaps:

Gap 1: Product Incoherence