AI report card · 2026-06-07 to 2026-09-18

B

gitzman

314 tasks, 74 deliverables, $5,420 of AI spend. Strongest: delegation. Most to gain: verification and specification.

What you built

74 things built with AI · $78k–$167k to hire them out to freelancers · built with $4,583 of your $5,420 AI spend

  1. EPA-per-play head and RL calibration feasibility review
    about 88 h of staff / architect work · in progress
    $8.8k–$18k freelance · $23k agencyAI $20
  2. Contractor outreach and reply pipeline
    about 104 h of senior engineer work · in progress
    $6.4k–$13k freelance · $21k agencyAI $150
  3. Agentic speculative decoding draft-model fine-tune benchmark
    about 80 h of senior engineer work · in progress
    $5.6k–$11k freelance · $17k agencyAI $331
  4. Difficulty-tiered BAML pipeline with plan review gate
    about 66 h of senior engineer work · shipped
    $4.5k–$8.9k freelance · $14k agencyAI $127
  5. Hosted token price and revenue tracker
    about 84 h of data analyst work · in progress
    $3.9k–$8.7k freelance · $15k agencyAI $118
  6. SWE-Gym specialization data pipeline
    about 45 h of senior engineer work · in progress
    $3.1k–$6.1k freelance · $9.4k agencyAI $132
  7. Cam Ward highlight video memorability pipeline
    about 36 h of staff / architect work · in progress
    $3.0k–$6.1k freelance · $8.5k agencyAI $197
  8. Inspection issue extraction consistency eval harness
    about 43 h of senior engineer work · in progress
    $2.7k–$5.4k freelance · $8.6k agencyAI $63

And 66 more in your dashboard.

Hours are estimated from the work in your sessions, for the role that would otherwise have done it. Freelance ranges are marketplace rates including the client fee; agency is a typical agency rate.

C

Verification 292 tasks

Does work finish with proof, not just a judgment?
22% verified · 43% completed

65 of 292 settled tasks finished with passing tests, a commit, or your explicit confirmation; you asked for verification on about 28% of tasks and tests ran in 3%. Published anchor: 15% of novice and 33% of expert sessions end in verified success.

A

Delegation 310 tasks

How much does the agent get done per prompt, and is big work right-sized?
15 actions per prompt

325 words of output per prompt; 84% of spend on hard, large tasks; 2% of large tasks abandoned. Published anchors: about 5 actions per prompt for novices, 12 for experts.

B

Hygiene and recovery 292 tasks

How much spend leaks, and does troubled work get rescued?
15% leaked · 34% of troubled tasks recovered

Abandoned $310, corrections $141, scheduled runs re-loading context $347; 2% of settled tasks abandoned. Published anchors: novices abandon 19% of troubled sessions, experts 5–7%.

C

Specification 310 tasks

How well do requests say what, where, and what done looks like?
1.0 of 3 specification · 5% corrections

Opening requests average 1.0 on a 0–3 scale (goal, scope, constraints, acceptance); 60 of 1131 prompts were corrections; 0% of tasks started with a plan.

C

Model discipline 314 tasks

Is routine work on frontier models, and is oversight spent where it matters?
22% of spend routable to claude-sonnet-5

94% of tasks ran auto-approved and you interrupted 3% of prompts. Experienced users auto-approve more and interrupt on exceptions rather than approving every step.

A

Leverage 314 tasks

How much value does a dollar of tokens return?
$25 back per $1

$137,175 of labor replaced from $5,420 of spend (1066 person-hours), plus $30,133 of work that would not have happened otherwise.

What would move the grade

  1. Verification. Ask for the check in the brief: tests must pass, show the diff, compare against the old output. Agents pay off most on work whose result can be verified.
  2. Specification. State the goal, the files in scope, the constraints, and how you'll judge it done. Ask for a plan first on anything large; experienced users do, and their results need fewer corrections.

Where the spend went

DevOps / infra · $1,153 · 21%Data analysis · $1,098 · 20%New feature · $886 · 16%Writing / content · $654 · 12%

Cut points come from published research where a number exists: actions per prompt, verified success and abandonment from Anthropic's study of 400,000 Claude Code sessions; oversight patterns from Anthropic's agent-autonomy research; specification and plan-first behavior from Sarkar's Cursor study; human-expert time as the unit of task size as in METR's time horizons. Elsewhere the cut point is a stated default, until a peer benchmark replaces it with percentiles. Tasks still in progress from the last 7 days are not counted as unfinished. A grade resting on fewer than 30 tasks is marked as a thin sample; small accounts have every task judged three times and merged so one flipped judgment cannot move a grade. Value is an estimate of labor replaced at role rates, not revenue.