# Token Maxxing Is Over and Here Is What Comes Next

## Executive summary

The talk argues that the current 'token maxxing' era for AI coding assistants is ending, necessitating a shift in focus from raw token spend to demonstrable Return on Investment (ROI). The speaker advocates for 'Sonnet maxxing'—using models that reason less and stop sooner—to keep human engineers actively involved in the development loop, which is critical for complex, rule-based systems like banking software. Key strategies include measuring gains in headcount rather than multipliers, implementing structured 'meta harnesses' (e.g., for Claude Code) to enforce a repeatable development lifecycle (refinement, specification, implementation, verification), and shifting left by making AI collaboration a multi-faceted, team-based refinement session.

## Key takeaways

- Sonnet Maxxing vs. Opus: The speaker suggests that models like Sonnet, which perform less reasoning and stop sooner, are better for keeping human developers 'in the loop.' This is crucial when dealing with complex, hard business rules, as it forces engineers to actively steer the model rather than passively accepting its output.
- Measuring AI ROI: Instead of focusing on lines of code or token spend, ROI should be measured by the customer value gained (e.g., was a 1% better home screen worth $20K in tokens?). The most effective metric is calculating the number of full-time engineers saved or gained by using AI.
- The Value of Team Refinement Sessions: To prevent engineers from falling into a siloed, one-on-one prompting pattern, the speaker recommends conducting 'refinement sessions' where the entire team discusses the feature idea with the AI agent present. This group setting allows peer pressure and collective knowledge to guide the conversation and prevent accepting the AI's default recommendations.
- The Meta Harness Approach: A structured 'meta harness' can be built on top of AI coding tools (like Claude Code) to enforce a standardized development process: refinement $\rightarrow$ specification $\rightarrow$ implementation $\rightarrow$ verification. This allows for structured metric capture and ensures optimal context window usage at each stage.

## Technical details

- Model Strategy: The speaker advocates for 'Sonnet maxxing,' preferring models that require less deep reasoning and stop sooner, thereby maintaining human control and involvement in the coding process.
- Codebase Analysis: A core engineering challenge is identifying undocumented or poorly understood sections of existing code. The speaker suggests a diagnostic test: opening a codebase and stopping when the function's purpose or business value is unclear, which is often surprisingly easy to locate.
- AI Prompting & Keywords: The outcome of AI-generated code can be significantly steered by specific keywords in the prompt. Using terms like 'scalable' or 'maintainable' forces the model into a different architectural direction than simple functional prompts.
- Meta Harness Implementation: A meta harness is an orchestration layer built on top of an LLM coding tool (e.g., Claude Code). It uses file watchers and structured stages (Refinement $\rightarrow$ Specification $\rightarrow$ Implementation $\rightarrow$ Verification) to capture metrics and ensure process adherence.

## Practical implications

- Shift focus from 'token spend' metrics to 'headcount saved' or 'customer value gained' when justifying AI investment to finance teams.
- Implement structured, multi-stage development processes (meta harnesses) to ensure code quality and track metrics at each phase (e.g., specification vs. implementation).
- Mandate team-based 'refinement sessions' for feature development to prevent individual engineers from working in isolation and accepting AI's default suggestions.
- When prompting AI, explicitly include architectural keywords ('scalable,' 'maintainable') to steer the model's output toward desired engineering principles.

## Topics

AI Coding, LLM Orchestration, Build Engineering, ROI Measurement, Software Development Lifecycle, Prompt Engineering, Pepijn Senders, HelloFresh, Epilogue, Sonnet, Opus, Claude Code

Source: https://www.youtube.com/watch?v=b84hZzCZzPM