The Golden Calf

Derek's Excel file is 847 megabytes. It contains 14 tabs, 2,300 rows, and a macro system he spent four months building. The file manages his AI prompts. Not the outputs. Not the data. The prompts. He has categorized them by domain, tone, output format, model version, and confidence score. He calls it "The Twelve Laws of the Melting Prompt Layer." He presented it at an all-hands meeting. People applauded.

Derek is building a cathedral on a foundation that is actively dissolving. Prompt engineering — the craft of carefully constructing inputs to language models to produce desired outputs — is a transitional skill mistaken for a permanent discipline. Every model update that improves instruction-following, every interface improvement that adds structured inputs, every agent framework that handles prompt construction automatically erodes the value of Derek's 847 megabytes.

The process has become a parasite. It consumes more organizational energy than the work it facilitates. Derek spends 22 hours per week maintaining, updating, and evangelizing his prompt system. The system generates approximately 4 hours per week of productivity gains for the team. The ratio is inverted, and nobody has run the numbers because the system looks impressive and "AI initiatives" are not questioned in the current climate.

Process parasitism has a rich history. The Six Sigma black belt certification industry generated $3.2 billion in annual revenue at its peak, creating a professional class whose primary output was the maintenance of the Six Sigma process itself. Agile coaching followed the same trajectory — a useful insight about iterative development calcified into a certification ecosystem, a conference circuit, and a job title that existed to perpetuate its own methodology.

The pattern is consistent: a genuinely useful technique emerges, early practitioners extract real value, the technique becomes institutionalized, the institution requires administrators, and the administrators become a cost center that justifies itself through the complexity of its own maintenance. The technique remains useful. The apparatus around it does not.

Prompt engineering is following this trajectory on an accelerated timeline. What took Six Sigma fifteen years — from useful methodology to bloated bureaucracy — is taking prompt engineering approximately eighteen months. The acceleration is driven by the speed of the underlying technology's improvement. The skill is being automated faster than the profession can establish itself.

LinkedIn data shows a 340% increase in profiles listing "Prompt Engineer" or "Prompt Engineering" as a primary skill between January 2023 and January 2025. Job postings with "prompt engineer" in the title peaked in Q3 2024 and have declined 28% through Q4 2025. The curves are crossing.

Organizations that hired dedicated prompt engineering teams in 2023 and 2024 are reporting a consistent pattern: initial productivity gains followed by increasing overhead as the prompt management systems grow in complexity. A financial services firm in Chicago built a prompt library of 4,200 templates. Maintaining the library now requires 1.5 full-time employees. The library's usage logs show that 73% of employees bypass it entirely, preferring to write their own prompts or use the model's default interface.

Derek's Excel file will not be maintained by his successor. It will be opened once, briefly examined, and abandoned in favor of whatever interface the next model version provides natively. The 847 megabytes will sit on a shared drive, untouched, a monument to a transitional moment that mistook itself for a destination.

The process-as-parasite pattern for prompt engineering is in early stage but accelerating. The compounding gap between current prompt engineering practices and improving model interfaces will reach a critical threshold within 6 to 12 months, at which point the elaborate prompt management systems being built today will become not just unnecessary but actively counterproductive — adding latency and rigidity to workflows that the base models handle natively.

The prompt engineering profession will not disappear entirely. A residual function will persist in specialized applications — safety-critical systems, highly regulated outputs, and domains where deterministic behavior is required. This residual function will employ approximately 5% to 10% of the current prompt engineering workforce, and it will look nothing like Derek's Excel file. It will look like software engineering, because that is what it always was.

The golden calf is 847 megabytes of reverence for a tool that is already outgrowing the need to be worshipped. Derek will update his Excel file tomorrow. The model update that makes it obsolete will ship without a changelog entry that mentions him.

Read the full pattern analysis for a deeper understanding of the forces driving this assessment.