Shift from "Prompt Writing" to Attention Budget Management

Shift from "Prompt Writing" to Attention Budget Management

This course begins with the tool PMs use most frequently when interacting with AI: the prompt. The goal is not to collect clever phrasing tricks, but to reset your mental model - because a flawed understanding of how models operate will lead you to spend effort polishing details that have zero impact on output quality.

Running Example: The Product Manager of FreelanceFlow - a financial management and budgeting app tailored for freelancers. The PM needs AI to summarize customer feedback on the cash flow tracking feature, and to transform raw git commits into user-facing release notes.

Allocate Finite Attention Budget Instead of Maximizing Prompt Length

The traditional approach treats prompting like crafting a magic spell - assuming phrases like "think deeply and thoroughly" will automatically yield smarter answers. In reality, governed by the Attention Mechanism (covered in AI Literacy Lesson 4), an AI model operates with a finite Attention Budget per inference call. A PM's job is not to maximize text length, but to curate the highest-signal token set to feed that budget.

The direct implication: longer prompts are neither safer nor higher quality. Stuffing extra background "just in case," repeating an instruction three times "for emphasis," or listing every imagined edge case directly dilutes the Signal-to-Noise Ratio. Empirical studies show that model reasoning accuracy degrades as context length expands, even when far below the physical context window limit.

Comparative Example: The FreelanceFlow PM needs to summarize customer feedback regarding the cash flow tracking feature:

• Context-Stuffing Approach (Maximize Context): "You are a world-class principal product analyst with 15 years of experience at top tech enterprises. Please read extremely carefully, thoroughly, without missing any detail. Below is all raw user feedback. Provide a deep, comprehensive analysis..."
• Curated-Context Approach: "Summarize the 3 most critical insights from the user feedback below. Each insight must include 1 key finding and exactly 1 verbatim quote. Prioritize issues reported by multiple users over isolated edge complaints."

The second version is significantly shorter while delivering exact execution parameters: prioritization criteria (frequency of occurrence) and output structure (finding paired with quote). The first is purely decorative fluff.

Analogy: Writing a prompt is like writing a task brief for a brilliant new colleague who interprets everything literally - the more concise and objective the criteria, the more accurate their delivery; a verbose brief loaded with emotional adjectives forces them to guess on critical decisions.

Exercise 1.1: Below is a prompt a PM used to convert raw git commits into release notes for FreelanceFlow:

"You are an outstanding content marketing expert with years of experience writing release notes for top tech products. Write beautifully, professionally, and engagingly; do not write lazily. Here are the commits: (1) fix: negative balances not showing in red on dashboard, (2) feat: add CSV export button for monthly spending reports, (3) fix: crash when user enters comma instead of decimal point, (4) chore: upgrade internal charting library (no user impact). Write outstanding release notes, ensuring nothing is missed, as detailed as possible."

Rewrite this prompt using Curated Context principles: eliminate decorative fluff, retain core data, and introduce explicit exclusion rules for internal commits that do not impact end users.