Identify mis-delegation
What you'll learn
Mis-delegation is asking the LLM to do something it's structurally bad at. The fix is architectural — a tool, decomposition, or a different system — not a better prompt.
- Mis-delegation = task structurally bad for an LLM. Architecture problem, not prompt problem.
- Five canonical classes: exact arithmetic, fresh/real-time data, authoritative facts, large deterministic transforms, irreversible high-stakes calls.
- Diagnostic signature: prompt iteration plateau. If quality won't move with better prompts, the task is wrong.
Mis-delegation is when work has been handed to the LLM that the LLM is structurally bad at — and no amount of prompt engineering will rescue it. The fix is always architectural: add a tool, change the surface, decompose into a smaller delegatable piece, or hand the task to a deterministic system. Recognising mis-delegation early is high-leverage; it's the failure mode that wastes the most time when missed because the natural reflex (rewrite the prompt) doesn't help.
The classes of mis-delegation worth memorising: (1) Exact arithmetic / counting at scale — needs a calculator or code execution. (2) Real-time / fresh data — needs retrieval, search, or an API. (3) Authoritative factual recall outside training distribution — needs grounded retrieval. (4) Deterministic transformations on large structured data — usually a script or SQL query, not an LLM. (5) Tasks where the cost of a wrong answer is catastrophic and there's no human checkpoint — the delegation itself is irresponsible regardless of capability.
The pattern that distinguishes mis-delegation from a Description failure: you've tried better prompts, examples, role definitions, and structured outputs, and the failure rate hasn't moved. That signature — *prompt iteration plateau on a task that should objectively be solvable* — almost always means the task was wrong for the model, not for the prompt.
The fix vocabulary: 'add a tool' (calculator, code interpreter, search), 'change the surface' (move to a deterministic pipeline with the LLM only at one well-bounded step), 'decompose' (break the work into pieces the LLM is good at — summarisation, classification, reformatting — wrapped around the deterministic piece), or 'don't use an LLM' (some work is just a SQL query).
Key points
- Mis-delegation = task structurally bad for an LLM. Architecture problem, not prompt problem.
- Five canonical classes: exact arithmetic, fresh/real-time data, authoritative facts, large deterministic transforms, irreversible high-stakes calls.
- Diagnostic signature: prompt iteration plateau. If quality won't move with better prompts, the task is wrong.
- Fix vocabulary: add a tool, decompose, change the surface, or don't use an LLM.
Examples
Mis-delegation. SUM-by-group is a SQL/script job. The LLM at the front of the pipeline is fine for choosing the grouping, but the actual aggregation should run in code. Fix: have the LLM emit the SQL, run it, and feed the results back.
Mis-delegation without a search tool. Adding 'and please be accurate' to the prompt does nothing. Fix: add a search/retrieval tool, then the LLM can synthesize over real results.
Pitfalls
- Burning hours rewriting the prompt for a task that's structurally undelegatable.
- Assuming a bigger / smarter model fixes mis-delegation. Sometimes it lifts the floor a bit, but the architectural problem remains.
- Conflating 'the model could in principle do this' with 'I should ask it to.' Capability ≠ delegatability.
Source notes: 00-academy-basics/notes/02-ai-fluency.md
Ask Claude
Build a prompt with this lesson + your question, copy it, and open the Claude Project in a new tab.