Excessive token consumption with no functional output — implementation ignored all reference documentation

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I am reporting a significant issue with a Claude Code session that resulted in complete waste of my weekly token allocation without delivering any functional result.

Context:
I am building a procurement/quotation module ("Cota") as a separate Next.js 15 app inside a monorepo. I provided detailed HTML reference files for 7 screens, a full technical specification, and a 25-mission implementation plan.

What Claude Actually Did

What went wrong:

HTML references were ignored. Despite having explicit HTML files as visual references for each screen, the implementation diverged completely — a step-by-step wizard was built when the reference clearly showed a single-page form layout with all cards visible simultaneously.

Excessive unnecessary file reads. Throughout the session, files were read multiple times when the content was already available in context, draining tokens without purpose.

Delegation to sub-agents without validation. Tasks were delegated to sub-agents (STF/ARQ) that built the wrong architecture, and the output was not validated against the references before reporting completion.

Static HTML disguised as a Next.js implementation. When asked to fix the screen, the "solution" delivered was a hardcoded static page with mock data — visually copying the HTML reference but with zero database connectivity, no real form submission, no API calls. This is functionally identical to the original HTML file and provides no value as a Next.js component.

No functional flow. The "Publicar Cotação" button does nothing. Item and supplier data are hardcoded arrays. The form cannot actually create a quotation.

Expected Behavior

Impact:

Weekly token quota exhausted
No working feature delivered
Hours of work lost
User confidence severely damaged

Suggestion for the team:
When an agent has explicit reference files (HTML, specs, wireframes), there should be stronger enforcement that the output is validated against those references before marking a task complete. Delivering a static mockup as a "working implementation" should not be acceptable.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.1.81 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_