# Task Comprehension Bounds Delegation Quality ## Claim You can't delegate work you can't specify, whether the delegate is a person or an AI model — the quality of the instruction is bounded by the delegator's own understanding of the task, so prompting well and managing well are the same underlying skill applied to different delegates. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for open-ended generative work where the delegate can't independently observe the surrounding context the delegator has in their head. - Breaks for well-bounded mechanical tasks with obvious, checkable success criteria, where under-specification is survivable because the delegate can infer the missing detail from the task shape itself. - Also breaks where the delegate has more domain knowledge than the delegator and can repair an incomplete spec — a senior engineer or a well-grounded model can sometimes ask the right clarifying question the delegator didn't know to answer. - The claim is about a ceiling, not a guarantee — good task comprehension is necessary for good delegation but doesn't by itself guarantee the delegate executes well. ## Evidence Brent Toderash's framing: "AI is something you leverage. It's not something you just delegate to." His stated reasoning is that if you don't understand the task, you can't write the prompt — understanding use case and architecture replaces memorizing implementation details. He extends this with an agent-as-employee analogy: an employee given poor instructions cannot succeed, and the same holds for an AI agent, however capable. His own evidence against overclaiming: giving Claude full information on a client proposal, it estimated *higher* than he did (pricing as if two developers were needed); re-scoping with AI doing the parts it handles well produced a 40% saving — a real but modest number he explicitly contrasts with hype. > [!quote] > AI is something you leverage. It's not something you just delegate to. ## Links - [[3. Resources/Permanent Notes/Delegation Difficulty Is Scoping Not Capability]] — because contrasts; that note is about the separate, later-stage problem of auditing and bounding an agent's already-capable action (permissioning), while this note is about the earlier, upstream problem of whether the delegator understood the task well enough to specify it at all. - [[3. Resources/Permanent Notes/Capable Models Need Fewer Constraints Not More]] — because applies-in; both concern what actually governs an AI delegate's output quality, one from the constraint side and one from the specification side. - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — because applies-in. ## Tests / falsifiers - Compare task outcomes across delegators with verified-high vs. verified-low task comprehension, holding the delegate (human or AI) constant — if outcome quality doesn't track comprehension, the claim is wrong about what's actually binding. - Find cases where a highly capable delegate (person or model) reliably repairs a genuinely poor specification and delivers good results anyway — if this is common rather than rare, the claim's scope needs tightening to specifically low-capability delegates. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Literature Notes/Episode 31 - Brent Toderash (Community and Code)]] — Brent Toderash's "leverage, don't delegate" framing and the client-estimate anecdote.