prd-driven-context-engineering

PRD-Led Context Engineering

Memory as Infrastructure — an operating system for building products with AI agents

Your AI partner is brilliant in one session and amnesiac by the next. This repository is the fix: a fork-ready methodology that turns documentation into a knowledge graph humans and AI navigate together — so the 50th session is smarter than the 1st.

Quick Start · The Idea · The Lifecycle · The Skills · Live Demo Views

⭐ If this changes how you build with AI, star the repo — stars put this method in front of the next team drowning in context drift.

The Problem: AI forgets. Teams drift.

Every era of software solved memory its own way — and broke it its own way:

• Waterfall had Static Memory. Write everything down upfront: certainty, at the cost of change.
• Agile moved fast and created Fragmented Memory. Knowledge scattered across tickets, wikis, and chats; shared understanding evaporated.
• AI-assisted building adds a third failure: Amnesiac Memory. The model that architected your system yesterday has never heard of it today.

The common mistake is treating these as tooling problems. They are memory problems.

PRD-Led Context Engineering builds Shared Memory: it treats AI as a team member, not a tool, and keeps documentation synchronized with code so humans and AI navigate the same truth.

The Idea: Memory as Infrastructure

This methodology comes from two converging experiences.

Leading human teams — alignment always followed the same pattern: rally around a single Source-of-Truth artifact and the team moves as one. Without it, even great talent drifts.

Partnering with AI — sometimes the model performs at a senior level, sometimes it hallucinates. The variable was never the model's intelligence. It was the Context Density provided: rich, structured context in; senior-level output out.

The convergence: documentation is not an afterthought. Documentation is the infrastructure of shared memory.

The Golden Rule: If it isn't part of the memory infrastructure, it isn't true.

So every durable decision gets a Unique ID (UJ-101, BR-004, API-045) in a Source-of-Truth file. That ID is a memory node with weight: when the AI references BR-004, it isn't guessing — it's retrieving a specific, validated decision you encoded. The linked network of IDs across files is the Knowledge Graph, and it lives in plain markdown, in your repo, under version control.

The 4 Pillars

1. Just-in-Time Context — IDs let you load only what a task needs. No repo-dumps, no noise, fewer hallucinations.
2. The Documentation Ecosystem — PRD ↔ EPICs ↔ SoT connected by links, skills, and hooks. When logic changes, both human and AI know.
3. Context Validation — context is measured like code. Too sparse, the AI drifts; too dense, it chokes. The repo scores itself (see Readiness Scoring).
4. Progressive Documentation — update in place, never copy. No PRD_v2.md, ever. One document, many versions, single current reality.

The Cognitive Shift

This changes how work is measured, not just how it's tooled:

What's in the box

Everything below ships in this repo, works offline, and forks in one click:

🧠 Feature: A Knowledge Graph in plain markdown

The pitch: long-term product memory with no database, no SaaS, no lock-in — just files with discipline.

The architecture is 3 + 1 + SoT + Temp, designed to manage Context Density for both human cognitive load and AI context windows:

1. Executive Functions — orient attention. Files load stable→volatile to maximize prompt-cache hits:
   • README.md — the Dashboard (where am I? what is active?)
   • PRD.md — the Strategy (why and what)
   • CLAUDE.md — the Physics (how the AI must behave)
2. Focus Memory — epics/: the only variable state. An EPIC frames one problem as one context window.
3. Long-Term Memory — SoT/SoT.*.md: the immutable facts. Business Rules (BR-), User Journeys (UJ-), API Contracts (API-), and 18 more ID types. Nothing duplicated; everything referenced by ID.
4. Short-Term Memory — temp/: the scratch pad. Files attach to the active EPIC and get harvested to SoT before the EPIC closes.

Just-in-Time Context: instead of dumping documentation into the context window, reference specific IDs (UJ-101, API-002). Fewer input tokens, deeper understanding.

📈 Feature: The Progressive PRD

The pitch: the "One-Shot" — asking AI to build the whole app at once — produces generic code and rapid drift. The Progressive PRD makes that impossible by design.

PRD.md is a gated workflow, not a document. The AI focuses on one stage at a time, and no stage advances until its Definition of Done is met:

Why gates work: constrained focus prevents the AI from guessing the architecture before it understands the users; deep focus produces meaningful IDs; the result is not just a working product but a desirable one.

The paradox that makes it practical: gates provide focus; the ecosystem provides agility. Because documentation is modular and interlocked, you can revisit any stage just-in-time — customer feedback during Build doesn't restart the plan, it updates the BR- rules and lets hooks propagate the change.

🛠 Feature: 47 skills, one for every decision

The pitch: the lifecycle isn't advice — it's executable. Every stage ships with skills that know what to consume, what IDs to produce, and which gate they feed.

• 41 stage skills (prd-v01-* → prd-v10-*): problem framing, competitive landscape, pricing, persona definition, journey mapping, risk discovery, architecture design, epic scoping, test planning, release planning, GTM strategy, case studies…
• 6 methodology operators (ghm-*): gate checks, SoT building, ID registration, insight harvesting, status sync.
• Named frameworks, encoded: April Dunford positioning, Alex Hormozi offer construction, Owned/Rented/Borrowed channel allocation, Geoffrey Moore chasm crossing, Teresa Torres continuous discovery, Rob Fitzpatrick Mom Test interviews.
• Three depth modes — quick (founder gut-check, <15 min), standard (default), deep (investor-ready, with assumption logs) — so the method scales from solo founder to team.

Every skill emits Consumes / Produces sections in SoT IDs, which is what keeps the knowledge graph connected as you move through stages.