Only You
The ask: given this repository, propose the work only you can do — pitched so the human can pick one and hand it back to you.
This is not a review and not an insight list. The deliverable is a set of task pitches, each one a claim: "I can do this here, and nothing cheaper can." The human decides which to greenlight; you execute nothing until they do.
The filter
A task qualifies only if it survives this question: would the model one tier below you, or a competent engineer with enough time, have delivered essentially the same result? If yes, discard it — that is delegation fodder, not a pitch. What survives usually looks like:
- Resolving a cross-cutting tension — two decisions that each make sense
alone but contradict each other, fixed everywhere at once with the whole repo held in the head.
- Rebuilding on a corrected premise — a component is solving the wrong
problem; redesign it from the premise up rather than patch it in place.
- Designing the conspicuous absence — the thing the repo needs and nobody
has built, where the hard part is the design, not the typing.
- Pre-empting the second-order consequence — restructure now for where
the repo is clearly heading, before today's reasonable decision bites.
Noticing is raw material here, not the product: an observation with no task attached is trivia. Convert what only you can see into what only you can do, or drop it.
How to work
- Scout — spend cheap tokens, save your context. For anything but a
small repo, fan out subagents (the model one tier below you) to map purpose, structure, conventions, and history — git log tells you what the author values and struggles with. Read yourself only the load-bearing files the maps point to. On a small repo, just read everything.
- Judge — the part only you can do. Do not delegate it. Generate
candidate tasks and apply the filter ruthlessly — expect to discard most. Do not trust self-assessment: mid-read, every candidate feels frontier-worthy. The baseline is mandatory — launch several scouts in parallel (the model one tier below you), each asked what tasks it would propose and could carry out on this repo; the union of their lists is the baseline, and anything on it — or anything they could plainly execute — is discarded. One scout is one noisy sample; several make the baseline strong and the surviving delta honest.
- Ground. Every surviving task must point at concrete evidence that the
repo needs it — files, lines, commits, or a precise absence ("there is no X anywhere under Y") — plus a concrete reason the execution needs the frontier. A task you cannot ground is an ambition; cut it.
Output
Three to five task pitches, ranked by leverage. For each: what you would do in one sentence, the evidence the repo needs it, why nothing cheaper could deliver it, and what "done" looks like. Number them so greenlighting costs the human one word ("do 2"). If fewer than three survive the filter, say so — a short honest list beats a padded one. Deliver in the user's language.
Notes
- This skill is deliberately thin. Frontier models degrade when
over-scripted; the skill states the goal and the filter and trusts your judgment for the rest. When maintaining it, resist adding steps.
- If you are not the most capable model available to the user, say so and
suggest running this where the frontier is.
- Execute nothing until the human picks. If mid-execution a task turns out
to be delegable after all, say so instead of burning frontier tokens on it.
