You've worked with AI long enough to notice something: some sessions feel sharp and useful, others feel like starting over. The AI asks for context you gave it last week. It contradicts a constraint you set in the last conversation. It produces something technically correct but completely wrong for your project.
The problem isn't the AI. The problem is that nothing carries between sessions. You have a project — with conventions, decisions, constraints, history — and the AI has none of it. Every new conversation, it's starting cold.
Scaffolding is the infrastructure that fixes this. Not by making AI smarter — by giving it what it needs to stay aligned. CLAUDE.md files, context documents, rules files, handoff protocols. The structural layer that makes AI a reliable collaborator instead of a useful-but-forgetful tool.
This module is about recognizing what that layer looks like when it's absent — and what happens to your work when it is.
A software team has been using AI coding tools for six months. They're not beginners. They use Claude Code daily. They have a monorepo with three services, a test suite, and a deployment pipeline. They ship.
But something is wrong.
Every few weeks, a developer spends the better part of a morning re-explaining the project to the AI. The conventions, the constraints, why certain patterns are banned, which service owns which data. The AI doesn't know. The developer types it out again. The AI follows it for the session. Next session: gone.
Last month, an AI session introduced a pattern they'd explicitly banned in a code review six months ago. The pattern was subtle enough that it passed review. It made it to staging. A senior engineer caught it. The conversation in Slack was uncomfortable.
This month, a new developer joined. Their first AI-assisted PR included three separate deviations from project conventions — not because they weren't paying attention, but because the AI had no way to know those conventions existed. The team's CLAUDE.md was a three-line file written during setup and never touched again. It said nothing useful.
The AI isn't broken. The scaffolding is. The team has no reliable way to tell the AI what the project is, what decisions have been made, or what boundaries to stay within. Every session starts with noise. Every new developer starts cold. The AI is powerful and uninstructed.
This module teaches you to recognize what that looks like — and what to do about it.
Scaffolding is not a single file. It's a system of three layers, each doing a different job. When any layer is missing, AI work degrades in a specific, predictable way.
This is what tells the AI what it's working on: the project's purpose, conventions, constraints, banned patterns, architectural decisions, and vocabulary. Without this layer, the AI treats every session as a blank slate. It invents conventions rather than following yours. It can't distinguish between patterns that are common in the world and patterns that are correct for this specific project.
The primary file for this layer is CLAUDE.md — but it doesn't work as a static document written once. It's a living record of decisions, updated when decisions are made.
This is what tells the AI what it needs to see before it can work. Context engineering is the discipline of deciding what code, documentation, and prior decisions should be visible to AI before each session — and maintaining that as the project grows. Without this layer, the AI works in a window that's too narrow. It solves local problems without knowing global constraints.
This is what governs how work moves between AI sessions and human review. Handoff protocols, session boundaries, escalation triggers — the operational layer that makes AI work auditable and reversible. Without this layer, mistakes accumulate silently. AI output enters the codebase without checkpoints. When something goes wrong, there's no record of what the AI was told or what decisions it made.
The GOVERN function requires organizations to establish policies, processes, and accountability structures for AI use before problems occur. For project scaffolding, this means: documented roles (who can approve AI output), documented constraints (what AI is and isn't authorized to change), and a defined process for how AI outputs enter production. If none of these exist in writing, GOVERN requirements are unmet — even if the AI is producing good output.
Users of AI systems (which includes developers using AI coding tools) have obligations to monitor AI outputs and maintain human oversight. Scaffolding is the mechanism that makes human oversight practical. Without context files and handoff protocols, oversight becomes inspection theater — humans technically reviewing but lacking the context to catch AI drift.
Systems evaluation — the skill of identifying measures or indicators of system performance and matching them to goals — applies directly to scaffolding. A well-scaffolded project makes it possible to measure whether AI is staying on track. An unscaffolded project makes this impossible: there's nothing to compare output against.
You can diagnose a scaffolding failure without seeing the CLAUDE.md. The symptoms show up in the work. Three things to look for:
How much time does a developer spend at the start of each AI session providing context the AI should already have? If the answer is more than two minutes, scaffolding is missing. Every minute of context-setting is a minute of work that should be in a file, not in a chat window.
How often do AI outputs introduce patterns, naming conventions, or architectural decisions that contradict the project's established choices? Once per week means Layer 1 (project identity) is absent or outdated. Every drift that isn't caught is a liability that compounds.
When a new developer joins and uses AI assistance, how many of their early PRs contain deviations from project conventions? If the answer is more than zero, Layer 1 isn't doing its job. The CLAUDE.md should encode enough project identity that a developer with no prior context can contribute conventionally from day one.
You'll apply all three of these diagnostic signals in the lab — working through a project description and identifying exactly where each layer is missing or broken.