forage

forage

!checklist init ${CLAUDE_SKILL_DIR} --force

A useful repository is a vein of value in a commons that, from the trending page, looks like nothing but noise and abandoned star-farms. forage is the lens you hold over the open-source world to find that vein before you sink an afternoon into the wrong one — whether you arrive with a sharp need or just wander in to look around. It is the first skill of the quarry suite: it owns discovery — bringing back the few candidates worth a closer look. 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 supply it, so the judgment is yours.

The one mental shift everything hangs on — the search syntax is no longer the skill. The human-era craft of finding repos was fluency: memorize the qualifiers, learn which awesome-list to grep, build the muscle for language:python stars:>1000 pushed:>2026-01-01. The agent already knows every qualifier and can run forty queries while you type one — so that fluency is commoditized, and clinging to it is optimizing the part the machine took over. What is left for you is the part it can't do: in targeted mode, saying precisely what you need (the sharper the spec, the better everything downstream — "a task queue" finds noise; "a lightweight Celery alternative with no broker dependency, still maintained in 2026" finds the answer); and in serendipitous mode, the taste to feel which find is worth a second look. The agent's job is to fan the seams out wide; yours is to aim them and to judge what comes back.

The agent is the means, not the judge. Here it has two jobs: the parallel miner (run every seam at once, targeted) and the wanderer (range the high-signal surfaces, serendipitous). It is not the arbiter of what's good. Its gradient points at handing you something that looks like an answer, and in this domain that gradient is dangerous in three specific ways: it trusts the most gameable signals — a big star count and a polished README — exactly the two an abandoned project or an AI-generated star-farm wears best; it invents repositories, or asserts that a real one "supports X" without ever opening it; and it flattens everything to "relevant", returning ten more of what you already know when serendipity needed the one orthogonal thing. So two disciplines hold from the first query: every candidate carries its provenance — which seam found it, which query — and no repo enters the list that the agent did not actually pull from a source; and stars and the README are treated as claims to be corroborated later by touchstone, never as proof of quality here.

What you cannot delegate — two judgments. The pipeline parallelizes the labor, but two points are yours and stay yours: (1) the need-spec (targeted) or the adjacency dial (serendipitous) — what you are hunting, and how far from your usual ground to roam, decides what can even be found; hand the agent a vague need and you have automated a thorough search of the wrong commons. And (2) the "worth a look" verdict — the 眼前一亮, the this one's interesting. The agent can rank by stars and recency; it cannot feel the pull that makes a project worth your weekend. Outsource that and you have a tidy list nobody actually wanted to open.

What "done" looks like — a captured shortlist, not an open tab graveyard. Foraging is over when you hold a small, categorized shortlist — each survivor with its provenance and a one-line reason it's there — recorded somewhere durable (a star-list, a notes file), not twenty browser tabs you'll close unread. Targeted foraging ends with a handful of genuine candidates framed for touchstone. Serendipitous foraging ends with the surprises captured cheaply into an "interesting, not sure why yet" list — because the value of a wander rarely lands the day you find it, but three months later when a problem arrives and you think "wait, I starred something for this." The terminus is the captured shortlist, not the number of repos you skimmed.