plan-researchlisted

Answer a research question with a corroborated, ranked, cited TLDR instead of a single-source guess, for any topic or project. Decomposes the question into sub-questions, fans out parallel agents that search the web AND the codebase and production data, then adversarially verifies every load-bearing claim (attempt to refute it, require a second independent non-conflicted source, discount sources by their stake). Ranks the surviving options best-to-worst for the actual use case, corrects the asker where the evidence says so, and hands back a human TLDR with links, never word-vomit. Tracks tasks and sub-tasks as they arise, sends research questions to spin-off agents and reserves real decisions for the user, and self-improves at the end. Use for "research X", "deep dive on X", "what is the best or most popular way to X", "compare these options", "is this claim true", "corroborate this", "what do people actually do for X", or before a decision that turns on facts you do not already hold.
skowalskidev/dotclaude · ★ 2 · Data & Documents · score 78

Install: claude install-skill skowalskidev/dotclaude

# Research → ranked, corroborated, cited TLDR The failure this prevents: research that is a single blog post restated, a confident answer from memory, a vendor's recommendation laundered into a conclusion, or three pages of word-vomit with no ranked answer. The output is a decision-grade TLDR: the best option first, why, the tradeoffs, and a link behind every load-bearing claim. This is the generic engine over four catalogs you READ rather than restate: `references/research.md` (the source-credibility method), `references/parallelization.md` (the fan-out), `references/tldr-report-formats.md` + `rules/copy-quality.md` + `rules/communication.md` (the output and where questions go). Open `references/research.md` before the first search. ## Step 1 · Frame the question and split it DO state the real question in one line and name the DECISION it feeds (what changes based on the answer). DO split it into sub-questions that do not overlap; each becomes one parallel finder. DO separate FACTS from DECISIONS. A fact ("does Stripe support single-use customer-bound codes?") is for an agent to find; a decision ("do we bill on it?") is for the user. Never ask the user a fact you can look up, and never decide for the user a call that is theirs. ## Step 2 · Track it DO write the angles and sub-tasks into a tracker (`.context/<topic>-research/TRACKER.md` or the Task tools) and add to it as sub-questions surface mid-run — the ones that surface are the load-bearing ones. Mechanics: `referenc