Generate & Read the Boondoggle Score
A complete Boondoggle Score has four structural elements and you should expect to see all four before you accept the output. First: every human step has a labeled supervisory capacity — Domain Verification, Ethical Judgment, Contextual Interpretation, Procedural Oversight, Adversarial Reasoning, or another capacity…
Generate & Read the Boondoggle Score
A complete Boondoggle Score has four structural elements and you should expect to see all four before you accept the output. First: every human step has a labeled supervisory capacity — Domain Verification, Ethical Judgment, Contextual Interpretation, Procedural Oversight, Adversarial Reasoning, or another capacity from the taxonomy. If a human step says 'review Claude's output' without labeling the capacity, the score is incomplete at that step. Second: every Claude step has a copy-pasteable prompt — not a description of what to prompt, but the actual prompt text. Third: every Claude step has
Every claim in this video was fact-checked against the source chapter and primary sources before rendering.
youtube.com/@NikBearBrown
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
[0:00] Ciao. This is Liam. Episode three of four. Run/clude. The boondoggle score is the answer to a specific question made explicit which work as the humans. Before you analyze your score, learn to read it. A score you cannot read is a document you cannot evaluate. This episode teaches you what a complete score contains, why human steps land where they do. There's a reason that is not intuition. And how to find the one step whose verification is most expensive. That step is your prompt deanker for episode 4. A complete boondoggle score has four structural elements and you should expect to see all four before you accept the output.
[0:33] First, every human step has a labeled supervisory capacity, domain verification, ethical judgment, contextual interpretation, procedural oversight, adversarial reasoning, or another capacity from the taxonomy. If a human step says review cla's output without labeling the capacity, the score is incomplete at that step. Second, every clawed step has a copastable prompt, not a description of what to prompt, but the actual prompt text. Third, every clawed step has a testable handoff condition. A testable handoff condition names what Claude delivers, the criterion for human acceptance, and the criterion for escalation. Claude delivers X, human accepts when Y, escalates when Z, a handoff condition that says the engineer checks the output
[1:13] Is not testable. Fourth, the score summary includes a distribution table showing the count of each supervisory capacity and two to three highest risk handoffs with a one-s sentence risk description for each. If your score is missing any of these four elements, that is a /claw prompt you need to repair before reading further. Here is the concept. The whole course turns on the solve verify asymmetry. Producing AI output is cheap. Verifying it is expensive. These are not symmetric. Claude can generate a candidate drug interaction list in seconds. Verifying that list, checking it against the patients actual medication history, the contraindication database, the attending physician's clinical judgment takes
[1:49] Domain expertise and time that the generation did not. That asymmetry is why human steps land where they do in the boondoggle score. Human steps are placed at verification points, not at generation points. The most expensive verification step in your score is not the most complex clawed step. It is the step where catching a wrong output requires the most domain knowledge. A wrong output that requires a pharmacist to catch is more expensive to verify than a wrong output that requires a format check. This matters for prompt D in episode 4. You are not looking for the hardest clawed task. You are looking for the most expensive human
[2:22] Verification. Keep that distinction in mind when you read the score. Reading your score, start with the distribution table. It tells you which supervisory capacity appears most often in your human steps and which appears least. That distribution reflects a property of your domain, not a property of the scoring rubric. A system with a high count of domain verification steps is a system where the humans doing oversight need deep domain knowledge at every handoff. A system with a high count of ethical judgment steps is a system where the human steps require judgment that cannot be fully specified in advance. Now find the highest risk handoffs. For each one, ask three questions. Is the
[2:58] Supervisory capacity labeled? Is the handoff condition testable? Does it say what claw delivers and what the human does with it? And does the acceptance criterion require domain knowledge to apply or could a non-expert run the check? A score that says human reviews Claude's output without answering those three questions is an incomplete score. Notice that incompleteness. The assignment wants you to notice it. That noticing is the skill. Verdict for generate and read the boondoggle score. Run /clawed with your completed SDDD. Complete score has four elements. Labeled supervisory capacities on every human step. Copa passable prompts on every clawed step. testable handoff conditions on every clawed step and a
[3:35] Score summary with distribution table and highest risk handoffs. The solve verify asymmetry production is cheap verification is expensive and that asymmetry not intuition is why human steps land where they do. Read the distribution table before the handoffs for each highest risk handoff check capacity labeled handoff testable acceptance criterion domain specific. Find the one step where verification is most expensive. Write it down. That is your prompt deanker. Your turn. Paste this prompt in agree with your completed STD. The prompt reads, "Here's my completed SDDD. Paste the full document. Run /claude and produce the full boondoggle score. For each claude step, include the copapastable prompt, the testable handoff condition in the format
[4:15] Cla delivers X, human accepts when Y escalates when Z, and the human supervisory capacity from the standard taxonomy for each human step. Label the supervisory capacity. Then add the score summary, the distribution table with capacity counts, and two to three highest risk handoffs each with a one-s sentence risk description. End quote. That is the full prompt. After you have the score, do one more thing. Find the one clawed step where the following human verification step needs the most domain knowledge to run correctly. That is the step where wrong output from Claude is hardest to catch without expertise. Write that step down by name. That is your prompt deankor for episode
[4:50] 4. You will need it. Generate and read the boondoggle score. Lay them in for bear.
