The Research Problem I Solved with a Slash Command
A Claude Code skill that turns URL pasting into three persistent artefacts: a public essay, an internal signal log, and roadmap backlog items.
You read about LangGraph on a Tuesday. You think: relevant to the agent work. You close the tab. By Thursday it's gone, not forgotten entirely, but not captured either. The signal decays into a vague sense that there's something worth looking at later.
That kind of loss doesn't feel expensive. It's quiet. Just gradual erosion of every interesting thread you encounter while building something else.
I built a Claude Code skill to close that gap: /research-capture.
What it does
Drop one or more URLs into the command. The skill fetches each page and synthesises the content into three things.
A public essay in /studio. Not a summary of the links, a synthesis: what the technology is, why it matters structurally, what it changes about how you'd build something.
An entry in docs/RESEARCH-SIGNALS.md. Internal, permanent log. Stack implications specific to the current agent infrastructure (Moirai, MakerSuite, Claude Code). Concrete integration points, not generic observations.
Rows in docs/ROADMAP.md Future section. Each suggestion from the signal log becomes a trackable backlog item with a link back to the source essay. The signal becomes part of the product backlog.
Three persistent artefacts from a single command.
Why this pattern matters
The IDE is already where you have context. You're mid-session, you find something useful, the natural instinct is to note it somewhere and get back to the task. The skill doesn't interrupt the session. Thirty seconds and you have material that compounds.
The essay is the public signal: here's what I'm tracking and why. The signal log is the private intelligence layer. The roadmap items are the commitment layer, where "this is interesting" becomes "this will eventually become work."
What the skill doesn't do is resolve whether to act on the signal now. That's intentional. Not everything worth capturing is worth building immediately. The skill preserves the option without forcing the decision.
The underlying system
The skill is global, in ~/.claude/skills/ rather than a specific repo. Any session in any project can invoke it. The output files accumulate in docs/ and content/thoughts/, building a long-term record of what I've been tracking and when.
Over time, RESEARCH-SIGNALS.md becomes a map of where the product is heading. The roadmap items link back to the evidence. The essays become the public record of how the thinking evolved.
That's the system. It pays off over months, not sessions.