Case Study

What 135 Days of Daily
Goose Usage Taught Me

Empirical findings from instrumented daily practice. Originally presented at Block's AI Tuesday series, June 2026.

Dakota Fabro (they/them)June 30, 20268 min read

135

Days of daily usage

$216

Total token spend

948

Impact points produced

14x

Efficiency multiplier

$216 over 135 days produced 948 impact points. That's 4.4 impact points per dollar. But the distribution is not uniform.

The Thesis

The Tool Didn't Get Smarter.
The Relationship Deepened.

Growth in human-AI collaboration isn't a linear increase in prompt sophistication. It's the emergence of compression, the rise of the floor, and the shift from explicit instruction to orchestrated delegation.

The same dollar produces 14x more value when the practitioner is thinking strategically. Low-DOK days yield 2.3 impact per 100k tokens. High-DOK days yield 42.8 impact per 100k tokens. The tool didn't change. The relationship to the tool changed. That's measurable.

The Core Finding

The same AI tool, the same token budget, the same practitioner. The only variable is the cognitive depth of the prompts. At DOK 3-4 (strategic and extended thinking), every dollar of token spend produces 14x the measurable impact compared to DOK 1-2 (recall and application).

The Data

A Four-Act Growth Story

PhaseDatesAvg DOKDOK 3+4Compression
1. OnboardingJan 28 - Feb 152.1417.4%3.3%
2. BuildingFeb 16 - Mar 152.3025.3%2.0%
3. PlateauMar 16 - Apr 302.1216.3%3.5%
4. DeepeningMay 1 - Jun 112.2929.3%9.9%
1

Onboarding

Learning the tools. Every interaction is explicit. DOK stays at Application level. No compression because every intent must be spelled out. The practitioner is building vocabulary with the agent.

2

Building

Framework creation burst. The peak (DOK 3.45) happens when the Three Dimensions paper is written. The tool is being pushed to its limits. Trust is building. This is the phase where the practitioner discovers what the tool can actually do.

3

Plateau

Execution-heavy phase. High-volume delegation days dominate. The raw DOK dips because the classifier can't see the orchestration happening above the prompt level. But compression begins to emerge (3.5%) as short directives start replacing long instructions.

4

Deepening

Strategic work returns with a new character. Compression triples (9.9%). DOK 3+4 hits 29.3%. The practitioner is now operating at Chamber/Orchestra trust level, issuing compressed directives that produce complex output. The gap between raw DOK and adjusted DOK widens because the cognitive complexity has moved from the prompt surface into the orchestration structure.

Findings

What the Data Reveals

1. Focused sessions produce deeper work (+10.5% DOK)

Session TypeDOKDOK 3+4
High-volume (>200 conversations)2.069.7%
Focused (≤50 conversations)2.2829.2%

This is the Orchestra Model in action. Delegation days look "low DOK" because the cognitive complexity lives in the orchestration setup, not in individual prompts. The classifier sees tool execution and scores it DOK 1-2, but the human decision to delegate is itself a DOK 3+ act.

2. Compression is the maturity signal

PhaseCompression %
Onboarding3.3%
Building2.0%
Plateau3.5%
Deepening9.9%

Compression means short directives orchestrating complex workflows. It's trust made visible in the data. When a practitioner says "proceed" and the agent executes a multi-step workflow, that single word carries the weight of all the context built over months of collaboration.

3. The floor has risen

67% of all sessions at DOK 2.1+
Longest streak above 2.1: 18 consecutive sessions
Dips below 2.0 are now rare - only on massive delegation days
The baseline DOK 1.77 would be impossible to reproduce today

4. Trust evolved from Duet to Chamber

Orchestra TierSessions%
Duet (Tier 2)314%
Ensemble (Tier 3)15%
Chamber (Tier 4)943%
Symphony (Tier 5)629%

Mean ADT: 4.0 (Chamber). Operating primarily at Chamber/Symphony level. Started at Duet. The Orchestra Model maps this progression: Solo → Duet → Ensemble → Chamber → Symphony → Virtuoso.

Implications

What This Means for Practitioners

Growth is non-linear

The plateau phase (Act 3) isn't stagnation. It's the period where delegation patterns are being internalized. The visible growth (Act 4) only happens because the invisible foundation was built during the plateau.

Measure depth, not volume

A 600-conversation day is not "better" than a 2-conversation day. The 2-conversation day where both prompts are DOK 4 produces more measurable impact than the 600-conversation delegation run. Volume is not a proxy for value.

Compression is the unlock

When you can say "proceed" and the agent executes a complex workflow correctly, that's not laziness. That's the culmination of months of context-building. The short prompt carries the weight of the entire collaboration history.

The tool is the constant

Goose didn't fundamentally change between January and June. The model improved, but the 14x efficiency gap exists within the same practitioner on the same tool. The variable is always the human's cognitive engagement.

Methodology

How This Was Measured

All measurement was performed by the rp-why Goose skill running locally against session history.

  • DOK classification is keyword-based on prompt text (conservative, tends to undercount orchestration-level work)
  • Compression detection requires 2+ signals: short prompt length, high response:prompt ratio, established session context
  • Adjusted DOK bumps compressed prompts by +1 level (capped at 4)
  • ADT is calculated from sub-agent delegation patterns (only available for sessions with sub-agent data)
  • Impact scoring uses a composite of deliverables shipped, PRs merged, and artifacts produced per session

Try It

Measure Your Own Practice

The rp-why skill is open source and installable in any Goose environment. Generate a baseline, track your growth, and see where your collaboration practice falls on the Three Dimensions.

# Install the skill

npx skills add https://github.com/block/agent-skills --skill rp-why

# Generate your baseline

/rp-why init

# Check your current session

/rp-why current

Related

Related Links