Untitled Lesson
Skill Matrix
UNLOCK NODES BY LEARNING NEW TAGS.
Why does providing an AI assistant with an entire large codebase as context not necessarily produce better results than providing a smaller, deliberately curated set of relevant files?
💻 Code Challenge | +75 XP
Write a well-structured project context file covering tech stack, test/lint commands, one key architectural convention, and one explicit thing to avoid, formatted for easy scanning.
An AI assistant kept generating code using a database access pattern the team abandoned 6 months ago, traced to an outdated project context file. Reorder the steps to fix this and prevent recurrence.
Task: Reorder the blocks in logical sequence to solve the problem.
A.D.A. Interface
Adaptive Didactic Assistant
Pascual Vila
Frontend Instructor // Code Syllabus
The Error //
Allowing a project context file to go stale, describing conventions or architecture the codebase has since abandoned
// Actively misleading if the codebase has since migrated:
"We use MongoDB for all data storage"
// Correct: kept current with the actual architecture
"We use PostgreSQL (migrated from MongoDB in Q2) via the repository pattern"The Solution //
A stale context file doesn't just fail to help — it actively misleads an AI assistant into generating code that matches an outdated pattern, which is worse than having no context file at all, since the assistant would otherwise have to infer conventions from the actual current code. Treat the context file as a maintained artifact, updated alongside significant architectural changes.
The Error //
Assuming providing more context (an entire large codebase) is always better than providing a smaller, curated, relevant subset
// Not necessarily better: an entire 200-file codebase pasted as context
// Better: deliberately curated to what's actually relevant
[the specific failing test, the function it tests, the exact error]The Solution //
Irrelevant context can dilute a model's attention on what actually matters for the specific task at hand, and every model has a finite context window meaning excessive irrelevant information can also crowd out genuinely relevant details. Deliberately curating context to what's actually relevant for the current task tends to produce better, more focused results.