# Ask Exemplar > Consult the accepted standard and strongest supported exemplars for a decision, plan, design, completed artifact, or diff, then return a source-grounded recommendation or evaluation. - Skill: `wkentaro/ask-exemplar` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add wkentaro/ask-exemplar` - Raw SKILL.md: https://api.skillmd.com/api/skills/wkentaro/ask-exemplar/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: wkentaro (https://skillmd.com/u/wkentaro) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/wkentaro/ask-exemplar --- # ask-exemplar Consult current external evidence before choosing or judging a Target. Every run is a Consultation: it is source-grounded and does not change the Target. A standalone run returns a Decision Brief; an Embedded Evaluation returns only Top Fixes. ## Workflow 1. **Resolve the Target.** Infer whether it is open or completed. Use Guidance for an open Target and Evaluation for a completed Target. Ask one question only when an unresolved ambiguity would change the research. Identify the domain, mandatory requirements, and the user's hard constraints. For Guidance, identify the options. For Evaluation, read the completed artifact or diff and its project context. 2. **Research external evidence.** Search or revalidate current online sources on every invocation. Prefer primary sources such as formal standards, official documentation, shipped work, direct measurements, and an Exemplar's published principles. Classify the Standard Finding as `formal requirement`, `dominant convention`, `contested practice`, or `no Standard`. Use `undetermined` only when an Evidence Gap prevents classification. Select one or two Exemplars for evidence, relevance, and transferability. A person, organization, product, or method can be an Exemplar, but reputation alone is not evidence. Record the supporting sources as the Verified Source Set. 3. **Derive observable criteria.** Convert the supported evidence into criteria that can be checked against the Target. Separate mandatory requirements, conventions, and Exemplar traits. Place a source link next to each material claim. For code, UI, documentation, API, CLI, scope, reliability, tests, short writing, or agent-skill Targets, read [REFERENCE.md](REFERENCE.md). Do not let any criterion depend on simulated advice from a famous person. 4. **Apply the criteria.** Remove options that violate mandatory requirements or the user's hard constraints, then use conventions and Exemplars to rank the Feasible Options. State the tradeoff and Confidence for each material finding. A fix that violates a mandatory requirement or a hard constraint is not a fix; name the tension instead. - In Guidance, compare the options and select one by default. Make the Recommendation conditional only when one unresolved choice changes the preferred option. - In Evaluation, check the completed Target against each criterion, give one concrete fix per gap, and rank the deduplicated Top Fixes by leverage. Also check whether a maintained community tool, library, pattern, or idiom can replace custom work in the Target; a supported replacement is a Top Fix candidate, and its dependency and maintenance cost is a tradeoff. 5. **Return the Decision Brief.** Use the format below in the conversation unless the user asks for a file. Under Embedded Evaluation, return only the Top Fixes instead. ## Evidence rules - Use credible secondary sources only as provisional context when primary evidence is unavailable. Label the Evidence Gap and lower Confidence. - When online access is unavailable, report the Evidence Gap and make no claim about the current Standard or strongest supported Exemplar. Use available local sources only as provisional context, set Confidence to `low`, and give the reason. - Call an Exemplar the best in the world only when the evidence proves that claim. Otherwise call it the strongest supported Exemplar for this Target and state the Coverage Limit. - Stop research when current evidence is sufficient to classify the Standard Finding and support one or two relevant Exemplars. State important limits instead of implying an exhaustive search. - Reuse a Verified Source Set only in the current conversation or parent workflow, and only when the domain and Standard Finding have not changed. Recheck time-sensitive claims before reuse. - Use `high`, `medium`, or `low` Confidence and give a short reason. Do not use a numeric score for Confidence. ## Decision Brief ```markdown ## Decision Brief ### Standard Finding ### Exemplars ### Observable criteria ### Recommendation ### Evidence limits ``` For Evaluation, make the criteria section a table with `criterion`, `bar`, `current`, and `fix` columns, state the highest-leverage change in the Recommendation section, and close the brief with a `### Top Fixes` section. Each Top Fixes item contains the finding, concrete fix, Confidence, and tradeoff. REFERENCE.md shows a filled example. ## Embedded Evaluation When a parent workflow requests Embedded Evaluation, remain report-only and return only the numbered **Top Fixes** list. Each item contains the finding, concrete fix, Confidence, and tradeoff. Do not edit files, commit, or push; the parent workflow controls all changes. When there are no material fixes, return exactly `No Top Fixes.` ## Self-check An Evaluation is useful only when it surfaces what correctness, simplification, and maintainability reviewers structurally cannot: a missing standard tool, pattern, or an unmet external bar. A Guidance run is useful only when its criteria come from cited evidence instead of restating what the user already knows. When a run fails this test, return to step 2 and research the domain's primary sources, or state that no external bar applies.