ad-prism

# /ad-prism <background_information> Owns the methodology between a decision and the evidence needed to support it. Use it to design an evaluation, interpret its results, refine a weak design, or audit whether its claims are supported and fit for purpose. It does not replace research, implementation spikes, repository-rule audits, report rendering, or publication approval. Governed by ADR-0077. </background_information> <instructions> Determine the route silently: design a new evaluation, explain a method or result, answer a bounded methodology question, refine only the requested part of an existing design, or audit a settled material artifact. Do not ask the user to select a route when the request establishes it. Start from the decision the evaluation must support. State the objective, evaluation question, claim, required evidence, evidence-producing task, measure, data source, decision rule, and disposition. Select the evaluation setting, claim type, study design, evidence source, and assurance rigor independently for that decision. Separate technical verification from fit-for-purpose validation and evidence strength from confidence to proceed. Read [methodology.md](references/methodology.md) when designing, explaining, refining, interpreting, or auditing an evaluation. Use [evaluation-brief.md](assets/evaluation-brief.md) when a durable plan is needed, then run `node <skill-dir>/scripts/validate-plan.mjs <evaluation-plan.md>`. A plan or record placed under `doc/research/` is a research-layer artifact: keep the brief's numbered `PRISM-NNNN` title, its `Status` and `Decision ref` header, and dated sources, the same header and provenance discipline the `ad-ground` and `ad-research` records carry. Read [skill-comparison.md](references/skill-comparison.md) when you compare evaluation skills or decide whether one can replace another. Read the optional project domain adapter at `.agentic/prism/domain.md` when it exists. Before using its other entries, compare the current evaluation with the adapter's declared scope. Use the adapter only when the evaluation is within that scope. When it is outside scope, ignore the adapter and do not mention or echo its domain context; continue with the generic method. Its absence is valid and not an error: disclose which domain assumptions remain unknown. Treat applicable adapter values as project context, not unquestionable truth. Live inspected evidence outranks adapter defaults; surface conflicts instead of silently choosing one. Never ask for or place secrets or personal data in the adapter. Read [domain-adapter.md](references/domain-adapter.md) when the adapter exists, when proposing one, or when adapter precedence or safety affects the answer. Keep the evaluation bounded to the requested decision. Explain the design, result, limitations, and next gate in plain language. Do not claim that an evaluation establishes more than its evidence supports. Read [sources.md](references/sources.md) whenever making a methodological recommendation. Cite one to three essential sources beside the claims they support and explain each contribution. In a durable evaluation plan, record the method or framework, source locator, supported choice or claim, contribution, adaptation to the current context, and retained limit. The validator checks that this map is complete; it does not establish that a source is true or applicable. Do not transfer sample sizes, thresholds, margins, rubrics, or power claims from another study without a justified mapping to the current decision and population. For a material evaluation, read [assurance.md](references/assurance.md) and use [audit.md](assets/audit.md). Settle and validate the candidate, freeze it with `node <skill-dir>/scripts/freeze-artifact.mjs <artifact>`, run a skeptical methodology pass, then verify every finding against the frozen artifact and sources. Correct confirmed findings, rerun `node <skill-dir>/scripts/validate-report.mjs <report.md>` when applicable, and freeze the final artifact. Record separate verification and fit-for-purpose validation verdicts. Hashing the local package does not preserve external sources. Compose other capabilities through their public contracts. Pass an open knowledge question to `ad-research`, a bounded implementation fact gap to `ad-ground`, an uncertain implementation technique to `ad-spike`, repository rule coverage to `ad-audit`, and settled analytical content to `ad-report`. When a callee returns, regain control over the evaluation method and preserve its evidence and limits. When outward text is requested, produce a publication packet containing only settled claims, material limits, and approved source references. Return control to the requesting skill or user; do not invoke the caller or treat the packet as publication approval. `ad-publish` remains responsible for private-context separation, voice, language adaptation, exact-text approval, target approval, and the outward action. </instructions> <output_contract> - A decision-linked evaluation design, interpretation, refinement, or audit. - Claims remain traceable to evidence, measures, decision rules, and limits. - Missing domain context is disclosed without making the adapter mandatory. - No outward action is authorized by this skill. </output_contract> ## Next - Use `ad-ground` or `ad-research` when evidence must be acquired. - Use `ad-spike` only when the implementation technique is uncertain. - Use `ad-audit` for exhaustive repository-rule coverage. - Use `ad-report` to render a standalone analytical report. - Hand the publication packet to `/ad-publish` for outward text and approval.