liveBook · Manning

Context blindness

As explained in Vibe Engineering

Context blindness is an AI-agent limitation caused by lacking relevant surrounding project information needed to perform a task correctly.

Context blindness is the failure of an AI coding tool to see the surrounding project information required for a correct implementation. An autocomplete suggestion can compile and pass tests while violating architectural decisions, cross-module type relationships, runtime assumptions, or conventions preserved as tribal knowledge. The result is design drift: locally reasonable patches that form a globally inconsistent system. section 5.1.3

A plausible endpoint with the wrong contract

For a payments endpoint, an autocomplete tool returns a dictionary and HTTP status code for success or failure. The project’s actual contract requires a typed Result[PaymentResponse], Result.expected_error for missing payments, and Result.success for valid responses. Both implementations look reasonable in isolation, but the generic version silently breaks downstream expectations. section 5.1.3

The limitation creates a tradeoff rather than a uniformly bad outcome:

  • Use context-blind autocomplete for boilerplate, exploratory prototypes, framework learning, and unfamiliar APIs, where generic patterns may be more useful than local conventions. section 5.1.3
  • Treat accepted suggestions as review work: fast generation shifts effort toward shallow, rapid evaluation and can weaken the ability to write solutions from scratch. section 5.1.3
  • Reduce blindness deliberately by supplying governed integrations and versioned procedures that package team-approved tools, checklists, scripts, and operating rules. section 3.3.6
  • Recognize the new cost: broader-context agents require approval of files, writes, and commands, so selective trust and risk-based supervision remain necessary. section 5.2.3

How Vibe Engineering covers it

How deep it goes
Explains causes, semantic consequences, task tradeoffs, and mitigations, but stops short of prescribing a complete verification workflow.
Examples
Uses Flask payment handlers, typed Result contracts, versioned agent skills, review checklists, and approval scripts.
What it assumes
Assumes familiarity with project conventions, architectural decisions, type relationships, runtime behavior, and code review.
What you can do after
Diagnose when generic AI output may violate hidden project contracts and choose lower-risk contexts or stronger supervision.
How it connects
Connects autocomplete, architectural decisions, tribal knowledge, design drift, agentic IDEs, selective trust, and supervision.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Vibe Engineering. The section links point to the parts of the book it draws on.