agentsclimarketplace
Sign in

Research driven development

Skill the-inconvenience-store/superbeads/skills/research-driven-development

Use when a decision or plan needs persistent repository, external, or mixed research evidence.From its SKILL.md

Install
npx -y skills add the-inconvenience-store/superbeads --skill research-driven-development

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.5 KB, as published. Nobody here has run it

Research-Driven Development

Apply the shared technical writing policy to research questions, reports, and user updates.

Research produces a durable evidence artifact before a decision or plan. The process may run inline or through bounded workers; the evidence contract is the same.

Announce at start: "I'm using the research-driven-development skill to investigate this topic."

Trigger / Non-trigger

Use this skill for comparisons, investigations, open questions, or unfamiliar topics whose answer will guide product, design, or implementation work.

Use a direct answer or ordinary repository inspection instead when the user explicitly wants a quick verbal answer, asks for one stable fact, or asks only what a named local file contains.

Iron law: research is complete only when docs/research/ contains the evidence artifact. A chat synthesis is an interim result.

Inputs

Establish:

  • the research question and the decision it informs;
  • known constraints and existing research;
  • the evidence mode; and
  • the expected depth: simple fact, comparison/decision, or open-ended.

Select exactly one mode:

  • repository-only when every load-bearing claim can be decided from this repository;
  • external-only when the answer is independent of the repository; or
  • mixed when the recommendation depends on both current local behavior and external facts.

After selecting, read research-modes.md for that mode's source, citation, and document contract. A mode change requires re-checking the evidence contract.

Steps

  1. Bound the question. If the decision, scope, or source class is unclear, ask only the questions needed to resolve it. Complete when the question, decision, boundaries, and mode are explicit.
  2. Open one research bead. Create and claim a task named Research: <topic>. Search bounded memories and docs/research/ for prior coverage. Complete when reusable evidence is identified or its absence is established.
  3. Plan evidence coverage. Split comparison/open-ended work into complementary, non-overlapping sub-questions. For repository current-state research, read question-planner-prompt.md: the controller that knows the request produces solution-neutral questions, then a fresh context receives only those questions and repository boundaries. Decision-informing repository observation does not stay inline; if the host cannot create a fresh context, persist the neutral questions and stop for a fresh-session handoff. Inline work is limited to decision-aware synthesis or a direct-answer non-trigger. When isolated workers would reduce clock time, read researcher-prompt.md, supply one bounded brief per worker, and cap a round at five workers. Complete when every load-bearing question has one owner and evidence class, and no repository-observer packet reveals proposed implementation details.
  4. Gather within the selected mode. Repository observers report current-state facts, contradictions, and unresolved evidence without recommending an implementation. Repository work uses LSP where available plus targeted rg, tests, history, and executable checks. External work starts broad, then narrows to primary or authoritative sources. Mixed work keeps blinded repository observation separate from decision-aware external research and synthesis. Complete when each load-bearing claim has decisive evidence or is marked unresolved.
  5. Synthesize once. Merge duplicates, test whether evidence entails each claim, resolve contradictions by authority and recency, assign confidence with a reason, and retain refuted claims when their removal affects the decision. Complete when findings answer every sub-question and recommendations follow from cited evidence.
  6. Close material gaps. Run at most one narrow follow-up round of one or two briefs for unresolved load-bearing claims. A remaining gap becomes an explicit limitation or blocker, never an invented source. Complete when the gap is resolved or its decision consequence is recorded.
  7. Write the artifact. Write docs/research/YYYY-MM-DD-<topic>.md or an existing matching category path. Use the mode's document contract and cite the current repository revision for repository evidence. Complete when a fresh reader can trace every load-bearing claim to evidence and understand the recommendation.
  8. Close the bead. File only evidence-backed follow-up work, then close the research bead with the one-line verdict. Complete when tracker state and the artifact agree.

Completion

Research is complete when all of these are true:

  • the selected mode's evidence requirements pass;
  • every load-bearing claim is cited, supported, and confidence-rated;
  • repository current-state findings remain distinguishable from later interpretation and recommendations;
  • contradictions, limitations, and refuted claims that affect the verdict are visible;
  • recommendations name an actor and action;
  • the research artifact records its date, bead, status, mode, and repository revision when applicable; and
  • the research bead is closed against the artifact's verdict.

Repository-only completion requires decisive path:line or command/test evidence and requires no external URL. External-only completion requires direct authoritative URLs for the claims they support. Mixed completion requires both evidence classes where the verdict depends on both.

Routing

  • Product behavior discovered or changed → superbeads:product-definition.
  • A solution decision is now ready → superbeads:brainstorming.
  • Requirements and design are already approved → superbeads:writing-plans.
  • Missing access or a user-owned decision → mark the bead blocked and state the exact recovery condition.

Conditional References

  • Read research-modes.md after mode selection; read only the selected branch plus the shared evidence shape.
  • Read question-planner-prompt.md before repository current-state research that informs a decision or plan.
  • Read researcher-prompt.md only when dispatching a research worker.
  • Read the repository's LSP/tool guidance only when repository evidence is in scope.

When this workflow closes after producing durable insights, read Durable Memory and apply it; otherwise do not load it.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.