target-scout — aim premium models at the right work

Understanding target-scout

In the overnight build system, I stumbled upon a useful tool called target-scout. Its purpose is clear: to direct premium models like Fable to the most impactful work. The key here is that target-scout uses cheaper models to score and map out where the high-leverage tasks are, ensuring that premium compute resources are only applied where they can make the biggest difference.

The target-scout documentation lives in C:\code\.claude\skills\target-scout\SKILL.md, and it outlines a two-pronged approach: Micro for within a single repository and Macro for decision-making across multiple repositories. If you're working within a specific repo, the default approach is Micro. However, if the question is broader, involving multiple projects, the Macro perspective is recommended.

Operating rules

The system operates on some strict guidelines. Scoring happens on two separate axes: Impact and Opportunity, which are then multiplied for a final score. Importantly, any score of 2 or less on either axis results in a discard, ensuring only the highest-impact tasks receive premium compute.

The tool itself runs on T1 (Sonnet) subagents, emphasizing the principle of using the cheapest reliable tier for this mapping process. This aligns with the policy of not using premium models like Opus or Fable for discovery—only for execution. If a task is identified as unlikely to benefit from machine intervention, it is skipped entirely, as outlined in references/03-escalation-and-skip.md.

Identifying constraints

Before initiating any scoring, it's crucial to understand the current business constraints. This is detailed in .claude/audits/.../FOUNDER-OPERATING-BRIEF.md and passive-brand-os/INDEX.md, which pinpoint what's considered P0 for the week. For instance, the current focus might be on closing the paid-conversion loop, a priority outlined in these documents.

For those working within TE-Code, it's recommended to utilize existing data from HOTSPOTS-AND-VALIDATION-FAN-OUT.md and CONTENT-FLYWHEEL-NODE-MAP to avoid redundant efforts. This ensures that the focus remains on the most relevant tasks, as per doctrine §2.6.

Systems pass and value loops

A significant part of the process involves mapping the flow of the system to identify the binding constraint. This involves running a Theory-of-Constraints pass, as described in references/08-systems-flow-and-constraints.md. The aim is to identify open or incomplete value loops—these are the areas that, if closed, could unlock significant value.

The process involves three steps: mapping the value loops, identifying the binding constraint (often at a seam), and sequencing the work to address these constraints. This results in a refined .claude/targeting/SYSTEM-CONSTRAINT-MAP.md, which serves as a guide to filter and prioritize the subsequent work.

By using target-scout, I can ensure that premium compute resources are applied precisely where they can have the greatest impact, driven by a clear understanding of constraints and system dynamics.