Article 8 of 12 · ARTHA Craft
Not theoretical. Production system. The institutional memory walked out the door.
This is not a theoretical problem.
This is a production system. Real users. Live data. A codebase built in the vibe coding era: fast, capable, and entirely undocumented. And one day a developer left for a better opportunity. Completely normal. Completely inevitable.
What wasn't normal was what happened next.
We sat down to continue the project, and realised we didn't know where we were.
The Crisis
Production system. Developer gone. No documentation.
Vibe coding era. Everything built fast, nothing recorded. The team knew the destination, but had completely lost the map of where they currently stood. A paralysing position for any codebase, let alone a live production system.
The Recovery
AI helped, but the hard way.
Asked the AI to do a deep codebase review. Reconstruct what existed. Surface what was partial. Map what had been built. It helped, genuinely. But it was painful, slow, and incomplete. The AI could read code. It could not reconstruct intent, half-finished decisions, or the architectural thinking that only existed in one person's head.
The Prevention
Asked AI: how do we make sure this never happens again?
The AI's answer was exactly what you'd expect from a system that understands its own limitations. Document context at the end of every session. Capture decisions. Write handoffs. Preserve state. This is standard AI best practice, because the AI knows better than anyone what happens when context disappears.
The Moment That Changed Everything
After the painful recovery exercise, after the AI had pieced together what it could from the codebase, I asked it a different question. Not "where are we." But: "How do we make sure this never happens again?"
The AI's answer was immediate, structured, and completely unsurprising to anyone who understands how AI actually works. Document your sessions. Capture decisions before they close. Write a handoff at the end of every working session that tells the next session exactly where to pick up.
This is not a novel insight. This is standard best practice, and the AI recommends it because it knows its own architecture. It knows every session starts from zero. It knows context disappears when the chat closes. It knows the only way to preserve state across sessions is to write it down before the session ends.
The AI didn't just help us recover from the crisis. It told us exactly how to prevent it. And the methodology that came out of that conversation has not changed since.
Seven sections. Each one answers a question the next session will need answered before it can do anything useful.
Session Date and Participants
When this happened. Who was involved. Which step was active. Simple. Permanent. Traceable. The baseline record that makes every handoff a dated entry in the project's history.
Step Status Table
Full snapshot of the step sequence at this moment. DONE, IN PROGRESS, PENDING, BLOCKED. All visible. The next session opens this and knows the full picture in thirty seconds without reading a single line of code.
Exact Stop Point
Not "we were working on the API." Surgical. "We stopped at: the E2 prompt recreated. E3 is next. File modified was X. Decision made was Y." The next agent picks up from the exact position, not an approximation of it.
Locked Decisions
Every design decision confirmed this session. The next agent reads these and treats them as final. Not up for debate. Not to be revisited without explicit approval. Locked: recorded here so they cannot be unintentionally undone.
Open Questions
Anything not resolved. The next agent surfaces these at the start of the session and waits for an answer before proceeding. It does not assume. It does not invent. It asks, because the question is recorded here and cannot be quietly skipped.
Standing Reminders
Permanent rules that carry across every session. Frozen modules. Patent considerations. Test coverage expectations. Written in every handoff until resolved, because these are the things that get forgotten in long projects and the forgetting is always expensive.
How to Open the Next Session
A paste-ready prompt. The architect pastes it verbatim into a new chat. No typing. No improvising. No "where were we?" The agent reads the handoff, reports position, surfaces open questions, waits for go-ahead. Every session starts from a known, documented state.
Before Closing Every Session: Ask This
"If I handed this file to a new agent tomorrow with zero prior context, could it pick up exactly where we stopped?"
A Developer Leaving
The project survives. A new developer reads the handoffs, the step sequence, the context seed. In one afternoon they know everything, without asking a single question.
Your Own Memory
You will not remember on Monday what you were thinking on Friday. The handoff does not rely on you remembering. It relies on you writing it down before you close the laptop.
Switching Tools or Models
You move from one AI model to another. The handoff travels with you. The new model picks up from exactly where the old one stopped, because the state lives in the document, not the session.
Any Gap in the Work
Holiday. Illness. Two weeks away. The project is exactly where you left it. No reconstruction. No "now where was I." The handoff is the bridge across any gap, of any length.