prospect

prospect

!checklist init ${CLAUDE_SKILL_DIR} --force

A research gap is a vein of value in a field that, from a distance, looks fully mined out. prospect is the lens you hold over a research area — at the very start of a project, the literature-review step — to find that vein before you sink months of compute, implementation, and writing into the wrong rock. It is the first skill of the inquiry suite: it owns step one of doing computational research — choosing what to work on, and proving it is worth working on. It audits (and guides you to run) a gated pipeline, and it will not advance past a GATE until the checklist tool clears it. That gate enforces order — each step done before the next — not the substance of the work inside it; the tool structures the discipline, it does not audit it, so the rigor is yours to supply.

The one mental shift everything hangs on — a gap is mined, not read. The human-era way was serendipity: read a survey and five to ten papers closely until an idea "emerges". That relied on luck and let one person run exactly one search thread at a time. The agent-era way is different in kind: you run a fixed set of extraction strategies in parallel, each blind to the others, to manufacture a candidate list of 10-20 gaps — and then you turn the agent against every candidate. A gap you read into existence is a hope; a gap that survived a deliberate attempt to prove it was already done is evidence. The whole pipeline is: bound the ground → mine candidates in parallel → attack and filter → rank by fundability → kill the survivors cheaply → land the three deliverables.

The agent is the means, not the oracle. Here it has exactly two jobs: the parallel miner (the manual labor — read 100 papers, build the matrix, cluster the limitations, reproduce a baseline) and the adversary (attack your gap until it breaks or holds). It is not the source of truth, and not — in this skill — the research object. Its gradient points at pleasing you: left soft, it will confirm your gap, invent a plausible citation, and read an abstract's marketing as an experimental result. So two disciplines are non-negotiable from the first search: the map is not the territory — every paper it surfaces has its title and core claim checked against the source before it enters your library — and every candidate is commanded adversarially, never asked to agreeably agree.

What you cannot delegate — three judgments. The pipeline parallelizes almost everything, but three points are yours and stay yours: (1) how to slice the dimensions — the matrix axes and the clustering themes decide what you can even see; (2) why a blank is blank — an empty cell is gold only if nobody thought of it, not if it is merely too hard, pointless, or already done under another name; (3) which gap to bet on — taste, after the filter and the cheap kill have done their work. Outsource these and you have automated your way to a confident bet on the wrong thing.

What "done" looks like — three deliverables, not "I read everything". The literature review is over when you hold: a one-page gap statement, a core-paper comparison table, and at least one or two baselines reproduced and running on your machine with numbers you personally verified. If you have those three, stop reading — you are done. If you don't, you are not done, no matter how many papers you have read. The most common failure of this step is the opposite of laziness: forever reading, never daring to commit. The terminus is the three artifacts, not the empty inbox of unread PDFs.

Speak the user's language. Most calls here are bets the user owns — which gap is worth their next year, whether a delta is "genuinely novel" or cosmetic, whether an assumption is "obviously unrealistic". Read their field fluency and gloss a term on first use (a seam, the adversarial existence search, reproduction arbitrage, the bitter lesson / time-discounting). A verdict the user can't evaluate is an opinion imposed, not a judgment shared — and the bet is theirs to place.