ad-research

name	ad-research
description	Research-to-conclusion methodology (WORKFLOW §17 + §4-5) — turn an open question ("should we adopt X?", "is approach A better than B?", "what is the state of the art on Y?") into an evidence-graded study at doc/research/NNNN-<slug>.md whose output is a defensible conclusion, not code. Runs the Evidence-Based loop (Ask, Acquire, Appraise, Apply, Assess) plus the scientific method (falsifiable hypothesis, experiment) when the question is measurable; reuses ad-ground's four-source pass for Acquire, grades each claim's evidence strength and the confidence to proceed (WORKFLOW §17), and graduates a binding conclusion into an ADR. Triggers on "research", "study", "investigate", "reach a conclusion", "should we adopt", "A vs B", "state of the art", "evaluate options", "literature review", "evidence for", "/ad-research". Distinct from ad-ground (records the path for an already-decided implementation) and ad-spike (validates one uncertain technique); routes to ad-spike when evidence is insufficient and to ad-adr when the conclusion binds.
summary	Research-to-conclusion — turn an open question into an evidence-graded study at doc/research/NNNN-<slug>.md via the Evidence-Based loop + WORKFLOW §17 grading + the scientific method when measurable. Reuses ad-ground for Acquire; graduates a binding conclusion to an ADR. WORKFLOW §17 + §4-5.

<background_information> Turns an open question into an evidence-graded conclusion — a study, not code. Implements the Evidence-Based loop (Ask → Acquire → Appraise → Apply → Assess) and the WORKFLOW §17 grade-of-evidence, adding the scientific method when the question is measurable. Acquire reuses ad-ground's four-source pass; this skill does not duplicate research mechanics. Output is a persistent study at doc/research/NNNN-.md, conclusion-first.

Use it for an open question whose answer is knowledge — "should we adopt X?", "is A better than B?", "what is the state of the art on Y?". Not for finding a path to implement an already-decided change (that is ad-ground), and not for validating one uncertain technique (that is ad-spike).

Codex auto-trigger on description keywords is less mature than Claude Code's. If auto-invocation does not fire on a research question, invoke this skill manually. </background_information>

Step 1 — frame the question (Ask) and determine NNNN. Turn the ask into an answerable question: the context, the candidate options, the criterion that decides between them, and the outcome that matters. State it in one sentence. If the ask is fuzzy, route to ad-grill-me first — a diluted question produces a diluted study. Name stakes times irreversibility for this question; this sets the §17 Axis-2 bar (how strong the evidence must be before the conclusion can read "proceed"). List doc/research/; NNNN is the next 4-digit number after the highest existing. If the directory does not exist, create it and start at 0001. Slug: kebab-case, ≤6 words, from the question.

Step 2 — hypothesis (when measurable). When the question can be settled by measurement, state a falsifiable hypothesis and the prediction that would confirm or refute it. When it cannot (a survey or state-of-the-art question), say so explicitly and skip — do not manufacture a hypothesis where there is nothing to measure.

Step 3 — Acquire (four-source pass). Run the four-source research pass — official docs, validated implementation references, in-repo patterns, git history — per ad-ground (WORKFLOW §4-5). Do not re-implement its mechanics; follow its Source A / B / C / D contract and cite as it does. This study is the durable evidence artifact for its Acquire pass; do not create a duplicate ground receipt. When a load-bearing claim can only be settled by measurement (evidence level E1 in §17), run an experiment or a staged spike (ad-spike, WORKFLOW §14) and bring the result back as evidence.

Step 4 — Appraise (grade the evidence, §17). Grade each load-bearing claim per WORKFLOW §17 Axis 1: seal it High / Medium / Low / Very-low, each with its provenance — citation, date, access method. A claim you cannot source cannot be sealed High. Where sources genuinely disagree, record the positions side by side — position A, position B, the evidence for each, and the conditions under which each holds — never a manufactured consensus.

Step 5 — conclude (Apply, §17 Axis 2). State the conclusion and its confidence per WORKFLOW §17 Axis 2: Strong (the answer holds — proceed), Conditional (holds with a named mitigation), or Insufficient / spike-first (the evidence does not yet support a conclusion; the gap is retirable by experiment — route to ad-spike). The bar for Strong scales with the stakes times irreversibility named in Step 1.

Step 6 — write the study. Write doc/research/NNNN-.md using the template below — conclusion first, evidence graded, sources dated. Status draft. Stop after writing; do not self-flip to concluded (that needs user review).

Step 7 — graduate (when the conclusion binds). After the study is reviewed, if its conclusion is a binding decision, graduate it: open an ADR (ad-adr) that references this study by path — the study holds the evidence, the ADR records the decision. If the conclusion feeds a pre-construction build, hand to ad-derisk (Front B) rather than building blind.

Interview: ask one question at a time for the framing choices (scope, stakes level, the candidate options) and the concluding verdict. Codex has no AskUserQuestion primitive; use inline numbered questions.

<output_contract> A single new file at doc/research/-.md, status draft. Conclusion first. Every load-bearing claim graded and sourced; genuine disagreement shown side by side, never collapsed. No invented values, no code. Documentation Discipline (WORKFLOW §2) applies — no emoji, business-context-first, definitions and decisions only — with one exemption: the study carries dates on its Status / Created fields and on every source in provenance, because dated evidence is the reproducibility primitive. No existing files are modified. </output_contract>

Next

• /ad-adr if the conclusion is a binding architectural decision — the ADR references this study by path.
• /ad-derisk if the conclusion feeds a pre-construction build — the study hands to Front B, which de-risks the build before implementation.
• /ad-spike if the conclusion is Insufficient / spike-first — run the experiment that settles the open claim.
• /ad-ground if the question turns out to be "how do I implement this?" rather than "what is true / what should I choose?".
• Flip Status to concluded after review; supersede with a new study when later evidence overturns it.