Research Decision

research-decision — sourced research that ends in a logged decision

This sub-skill stands up a research / decision project: a question, a set of candidates, a set of criteria, and a deadline — driven to a cited, defensible recommendation. It mirrors a prior research/decision project: a small fixed doc system, a berichte/ (reports) folder, a decision matrix scoring finalists on criteria, a portfolio of evidence, and the decision deliberately deferred to a date. Honesty is the whole product: every claim is sourced and marked verified vs. reasoned.

What this sub-skill is for

Answering a real decision with evidence: "which university / broker / framework / car / vendor / strategy", a feasibility study, a buy/build/skip call, a thorough literature scan. The output is a recommendation + scored matrix + a logged decision, not code. Composes with content-writing when the result is a long report, and with build-business when the decision is a business one.

Mandatory grill-questions (fold into the Definition of Ready)

Lock these before any research starts:

• The decision / question — one precise sentence: what is being decided, for whom, and why now.
• Candidates — the concrete options under comparison (and how open the field is — fixed shortlist vs. "find the candidates too").
• Criteria & weights — the dimensions that matter and their relative importance (cost, fit, risk, freedom, deadline pressure…). These become the matrix columns. Recommend deriving weights explicitly with the user, not assuming them.
• Deadline(s) — the real decision date and any per-candidate deadlines (e.g. application windows). Finalists often each have different cut-offs — capture every one.
• What "decided" looks like — the artifact + action that closes this: a logged choice, an application sent, a "defer until September" gate. Define the done state.
• Evidence bar — how strong must sources be (peer-reviewed / official / practitioner-ok)? What needs primary sources vs. acceptable reasoning?
• Reversibility & stakes — one-way vs. two-way door; what's the cost of being wrong. Drives how hard to grill and whether to convene council.
• Scope in/out — which candidates/criteria are explicitly excluded, to bound the research.

Project sub-agents to generate (into <project>/.claude/agents/)

• source-finder (delegate-by-default) — fans out across the web/literature to gather primary and high-quality sources per candidate/criterion; returns cited findings with URLs, never unsourced claims. Chains the deep-research skill.
• claim-verifier (delegate-by-default) — adversarially checks each load-bearing claim, marks it verified (with source) vs. reasoning/inference, flags contradictions and stale data, and downgrades anything it can't substantiate.
• decision-matrix-builder — maintains the criteria × candidates matrix, scores cells from the verified evidence, applies the weights, and surfaces the ranking + sensitivity (what flips it).
• Plus the kit defaults: reviewer (audits the report for unsourced or overconfident claims) and verifier (re-checks that conclusions actually follow from the logged evidence).

Tools / CLIs / MCP / skills needed

Check in environment-readiness; offer install, never auto-install:

• Research MCPs/tools — WebSearch/WebFetch; domain MCPs when relevant (PubMed, Consensus, Scholar Gateway, Clinical Trials for medical/academic; Semrush for market questions).
• Document tooling — xlsx for the scoring matrix as a real spreadsheet; pdf/docx to export the final report; pptx if a decision briefing deck is wanted.
• document-design + docrender (2026-07-29 addition) — the final recommendation report is exactly the "Entscheidungsvorlage" genre this skill is built for: its own layout (never a filled-in template), typeset in Typst, every page rendered to an image and checked with docrender review <file> before it's called done. Use it in place of a plain pdf export whenever the report is the actual deliverable someone reads and decides from, not an internal note.
• CHAIN these GLOBAL skills (the core of this type):
  • deep-research — multi-source, fact-checked, cited research per candidate/criterion.
  • council — convene the four-voice decision council on the real tradeoff before recommending, especially for one-way-door, high-stakes calls.
  • doc-coauthoring — structure the final recommendation/report.
  • verify — sanity-check that the recommendation follows from the matrix.

File / asset nudges (on top of the base set)

Beyond CLAUDE.md, PROJEKT_.md, TASKS.md, DONE.md, README, .claude/ — mirror a small fixed 4-file + reports layout:

• berichte/ (reports) — one cited write-up per candidate (and per major criterion), each marking verified vs. reasoned claims, with a sources list.
• decision-matrix.md (+ a decision-matrix.xlsx) — criteria × candidates, weights, scores, and the resulting ranking + sensitivity notes.
• candidates/ — a profile per option (facts, deadlines, pros/cons, open questions).
• DECISION_LOG.md — the running record: options considered, the chosen one (or the "defer to " gate), the reasoning, and what would change the call later.
• sources.md — the bibliography/URL list backing every claim.

Stack defaults & done-bar

Defaults: a small fixed Markdown doc system + a berichte/ folder + a weighted decision matrix (as both .md and .xlsx) + a DECISION_LOG.md; deep-research for gathering, council for the call, sources cited inline throughout. Defer the decision to the real deadline rather than forcing it early when the field is genuinely open. Done-bar (all true): every criterion is scored for every candidate from verified, cited evidence; a clear recommendation (or an explicit "decide on " gate) is written with its reasoning; the decision is recorded in DECISION_LOG.md; per-candidate deadlines are captured and none has silently lapsed; the report distinguishes verified facts from reasoning throughout.

Guardrails

• Cite or flag — never assert unsourced. Every load-bearing claim carries a source or is marked explicitly as reasoning/inference. No fabricated statistics, quotes, or citations, ever.
• Verified vs. reasoned is a visible distinction in every report — the reader must always know which is which.
• Don't fake confidence or completeness — surface gaps, contradictions, and stale data; say "I could not verify X" rather than smoothing it over.
• Honour every deadline — track per-candidate cut-offs; never let a window lapse unflagged.
• Separate evidence from recommendation — score on the evidence first, then recommend; convene council for high-stakes one-way-door calls instead of deciding solo.
• No emojis in any deliverable UI/output (user's standing rule).
• Commits under the user's own name only (Skryx-L-A); never add Claude as a co-author.