Article 6 of 12 · ARTHA Craft
Not a wiki. Not a README. Five documents with specific jobs.
At some point in every AI-assisted project, you hit the same wall.
You need to know what was built last week. You scroll back through chat history. You scan through sessions trying to find the decision. You piece together context from fragments of conversations that were never designed to be a record.
It works. Barely. Slowly. With no confidence that you found everything.
I spent weeks working that way. Then I stopped. Not because I found a better tool. Because I built a better system.
Before the system, let me explain why people try documentation and give up.
They build a wiki. It grows. Nobody maintains it. It gets stale. They stop trusting it. They go back to chat history. Or they build a README. It captures the starting state. The project evolves. The README doesn't. Six months later it describes a system that no longer exists.
Every document in this system has a job. A specific layer. A specific owner. A specific moment in the workflow where it is read, updated, and acted on. That is the difference between a wiki nobody trusts and a system that runs the project.
The Constitution
What the platform is and what it cannot be. Non-negotiables governing every decision system-wide. Module inventory. Architectural boundaries. The first thing any agent reads. Changes rarely, and only with an explicit decision, never a casual edit.
ContextSeed.MDThe Current State
What has been built recently. What is partially done. What is in progress right now. Not a history, a rolling snapshot. Keeps the constitution lean while capturing the living reality of where the product actually is today.
SecondaryContextFeed.MDThe Session Bridge
Written at the end of every session. Read at the start of the next one. Exact stop point. Locked decisions. Open questions. Standing reminders. And a paste-ready prompt: the architect pastes it verbatim into a new chat. No typing. No improvising. No blank slate.
CONTEXT-HANDOFF.MDThe Task Files
One document per feature. Background, design decisions, cross-layer dependencies, scope boundary, gap resolution, go-ahead prompt. The coding agent executes from this, not from memory, not from chat, not from an improvised brief. Written in the planning session, locked before the coding session opens.
STEP-XXXX-name.MDThe Registers
Three live documents: bugs, enhancements, deployment issues. Nothing gets fixed without a register entry first. Nothing gets forgotten because it lived only in a chat. Every item has an ID, severity, status, and traceable path to resolution. Reviewed at every sprint start.
BUGS.MD · ENHANCEMENTS.MD · DEPLOYMENT-ISSUES.MDEvery layer above is valuable on its own. But without one specific layer, they are five useful documents that don't quite form a system.
The Traceability Spine
The Step Sequence: REQCHAIN-STEP-SEQUENCE.MD
Every step ever designed lives here. Every step ever completed lives here. Status visible at a glance: done, in progress, pending design, blocked. History never lost: completed steps stay permanently. Nothing is ever removed.
But its job is bigger than a status tracker. Much bigger. The Step Sequence is the traceability spine of the entire methodology. Every other layer connects through it.
Planning Agent reads Step Sequence
Sees what has been built, what is planned, where the next feature fits. Designs from position, not from a blank slate.
Design work produces a Task File
Locked to a specific step ID. The ID is the traceability link: every decision traceable back to the step that created it.
Coding Agent executes against Task File
Builds from a locked brief. No design decisions to make. No gaps to fill. No invention required.
Step marked DONE. Handoff written.
History preserved. Context Handoff references the step ID. Next session opens, reads sequence, confirms position, picks up exactly.
Before the system
After the system
The speed does not come from the agent. [3] It comes from the system that feeds the agent. An agent with a clear, structured, complete brief moves faster and makes fewer mistakes than an agent trying to infer its brief from a chat history.