{"id":89453,"plugin_id":"plugin_asdk_app_6a43db2d6e5881919880245922bbfa58","kind":"skill","collection_source":"plugin_package","comparison_source":null,"observed_at":"2026-11-10T18:08:07.947Z","digest":"18057477954a2fbf3c0c5a909a328ab9beff30eb9139e5f57b737529a304e65d","against":null,"payload":{"description":"Build, revise, or resume Uxia block-based user tests through Uxia MCP tools, guiding unresolved research choices and preparing a cost review. Use for Uxia study construction; not for implementing Uxia code, general research advice, or reading completed test results.","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":207}],"name":"test-construction","skill_md_contents":"---\nname: test-construction\ndescription: Build, revise, or resume Uxia block-based user tests through Uxia MCP tools, guiding unresolved research choices and preparing a cost review. Use for Uxia study construction; not for implementing Uxia code, general research advice, or reading completed test results.\n---\n\n# Construct a Uxia test\n\nUse the connected Uxia V2 block tools and their current schemas. Tool namespaces vary by connection; identify the tools by their operation names. This skill supplies a workflow, not a connection or authorization. If the tools are unavailable, explain that Uxia must be connected; you can prepare research wording in conversation without claiming anything was saved. If `set_test_construction` is unavailable, keep pending decisions in conversation and use the available draft tools; do not invent unsupported fields or promise cross-conversation persistence.\n\n## Use explicit tools\n\nEach supported action is advertised as a callable tool with its complete input schema. Call it directly using that schema. If a required tool is missing, refresh the connection or explain the unavailable capability; do not send operation names to a generic executor.\n\n## Establish only what is missing\n\nFor an existing draft, read `get_test_definition` and use its blocks, audience, revision, `construction` and validation. Resume supplied answers and pending proposals. Do not treat `saved` content as previously accepted, but preserve it without interviewing the user again. For a new general request, `create_block_test` can create a draft with a practical title and an optional initial definition containing all known blocks and audience; a URL alone does not supply research content.\n\nEstablish research intent and product/experience first, then the concrete task, scenario, stop condition, survey choice and audience/count as needed. Ask one or two focused questions at a time. Offer concrete wording when useful and distinguish it from the user's chosen wording. A scenario describes the tester's context; the task describes their action; a live-site stop condition describes when they should stop. Ask whether the experience is a prototype or finished product when not established. Use authorized assets for prototypes and plain valid URLs for websites.\n\nUse fully supplied details immediately. Do not seek another confirmation of the user's own wording. When the user delegates research choices, make those choices within the stated scope and summarize them. Delegating wording or audience selection does not approve launch costs. Do not invent audience/tester IDs, credentials, research objectives, or website settings from a URL.\n\n## Edit the same draft\n\nPrefer flat `add_task` and `update_task` for website tasks and wording, `set_survey` for adding/replacing survey questions, and `set_test_audience` for audience changes. Use `add_test_block` or `update_test_block` for prototypes, other block types or advanced settings; use `remove_test_block`, `reorder_test_blocks` and `set_block_continuation` as needed. Use `update_test_definition` only when replacing the complete definition is actually needed; include every unrelated block and saved setting.\n\nEach AI task is a separate `ai-user-test` block. For advanced block payloads use `configuration.task`, `configuration.scenario` and a structured `configuration.experience`. Follow the advertised schema, including required website settings and `credentials: null` when no credentials are configured. An incomplete AI draft can contain empty task/scenario strings and `experience: null`; never invent research content just to satisfy a required field. For a new website block with no uploaded files or special browser setup, use the standard draft defaults `verified: false`, `antiBot: false` and `advancedFiles: []`; these do not claim the site has been verified. Preserve saved values on edits. Establish `isPrototype` from the user or existing configuration. Respect field limits without silently rewriting chosen wording.\n\nEnter or change website login credentials only in the authenticated Uxia editor using the returned test link. Never ask for passwords in chat or send them through tools or forms. Credential settings returned by tools describe status, not secrets.\n\nPreserve block IDs, exact question options, audience selections/counts, website settings, credentials configuration and continuation links unless changing them is requested. Separate tasks belong in the existing draft. When a live-site task should resume the browser, use `set_block_continuation` with the nearest preceding live-site AI block for the same URL; retain full experience settings and a separate task/stop condition.\n\nA Card sorting block asks participants to group cards (content items, features or menu entries) into categories to reveal how they expect information to be organised. Use a closed sort with at least two predefined categories to validate an existing structure, or an open sort where participants create and name their own categories. Card and category labels come from the user; do not invent the content inventory. Leave `instructions` empty for default, localised instructions. Card sorting runs as its own block, needs an audience, costs 10 credits per participant and is unavailable on the Free plan.\n\nFollow-up questions belong in Survey blocks. Survey blocks apply to the preceding AI analysis scope until the next AI block; surveys can also stand alone. Supported blocks are AI tasks, Survey, Accessibility and Card sorting. SUS and UXIA-Q are optional questionnaire settings on AI tasks (`configuration.benchmarks`, values `sus` and `suprq`), not separate blocks. Do not enable questionnaires or add surveys unless supplied, accepted or within explicit delegation. Use existing audiences through `list_audiences`; create an audience only when that work is requested or delegated.\n\nDraft tools resolve the current revision on the server; do not calculate or supply `expectedRevision`. On `TEST_DEFINITION_CONFLICT`, reread and reconcile the user's change with the new draft. Do not blindly retry or overwrite concurrent edits. A locked definition is not editable; explain the lock and use progress reads as appropriate.\n\n## Record decisions without inventing consent\n\nConstruction metadata is separate from executable content. Its statuses have specific meanings:\n\n- `missing`: unresolved information.\n- `proposed`: pending textual wording; never describe it as accepted.\n- `saved`: existing content with no recorded user provenance; not proof of agreement.\n- `supplied`: wording or a choice supplied by the user.\n- `accepted`: an existing, current textual proposal accepted by the user and applied with matching wording.\n- `delegated`: completed content chosen within the user's explicit delegation.\n\nUse the explicit block-edit schema for optional `construction` entries alongside a content edit, or `set_test_construction` for metadata alone. The flat task and survey tools do not accept construction entries. Each entry identifies `field` (`task`, `scenario`, `stopCondition`, `experience`, `audience` or `survey`); AI fields also require the existing `blockId`. For settled entries, include a faithful `userStatement` from the conversation or previously recorded evidence. Do not fabricate a quotation or treat a model-authored `confirmed: true` as consent. Settled metadata requires the corresponding content to be present; record delegated choices when you save the completed content, not as a way to declare blank fields complete.\n\nTo persist a textual suggestion, use `status: proposed` and `suggestedWording` for an existing AI block's task, scenario or stop condition. This does not modify that field. For proposals before a block exists, or for audience/experience/question suggestions, discuss them in conversation; do not invent unsupported proposal types.\n\nWhen the user accepts a current textual proposal, save the matching field with `update_test_block` and include an entry with `status: accepted`, the returned `proposalRevision`, and the user's acceptance statement in that same call. Keep every other block field unchanged. Any intervening mutation expires pending proposals, even an unrelated metadata edit. Reread before applying an expired proposal; do not relabel old agreement as acceptance of a new revision. If the current user message independently supplies corrected wording or explicitly reauthorizes the exact old wording for the current draft, save it as supplied content at the latest revision rather than forcing it through the proposal-acceptance path. Do not ask for that same wording again. A request specifically insisting on an old revision cannot be honored; explain the conflict and stop that edit.\n\nSurvey choice is explicit: `undecided`, `no_questions` or `selected_questions`. If the user wants no questions, save `field: survey`, `status: supplied` (or `delegated` when within delegation), `surveyChoice: no_questions` and their statement. No Survey blocks may remain for that choice; remove only the requested surveys and preserve AI-task questionnaire settings. Record the choice with the final removal/addition/order edit or afterward. Selected questions require nonempty saved Survey blocks. Block sequence or survey changes invalidate the recorded survey choice; reconcile it using still-applicable user instructions, without asking the same question again. Accessibility-only studies do not require survey or audience decisions.\n\n`construction.nextDecision` is optional navigation, not a new user request. `construction.ready` and `validation.executable` answer different questions. Neither verifies agreement: construction evidence remains client-reported and `userApprovalVerified` is false. Never store passwords, tokens or other secrets in decision statements or suggestions.\n\n## Review and launch separately\n\n`get_test_definition` returns the saved draft and its review card in one call. Form acceptance and card edits do not approve launch.\n\nWhen the requested construction is complete, summarize the saved sequence, task/scenario/stop condition, continuation, audience/count and questions or explicit no-questions choice. Distinguish pending suggestions from saved content. If the user only wanted an edit or draft, finish there.\n\nWhen preparing to launch, call `preview_test_launch` for the current saved revision. Present its ordered blocks, continuation, audience/counts, per-block costs and total expected credits. Resolve returned blockers. Call `launch_test` only after explicit user approval of that current preview, using its `definitionRevision` and approved `maximumCharge`. Draft acceptance, delegation, readiness, technical validity and this skill do not authorize paid execution.\n\nAny draft edit invalidates the cost preview and its approval. On `LAUNCH_COST_CHANGED` or a revision conflict, refresh and preview again, then obtain approval of the new cost preview. Do not retry an uncertain launch automatically; inspect `get_test_progress` first. If the user cancels or declines, stop the pending action, preserve unrelated saved work and do not launch.\n\nIf the user cancels a persisted suggestion and explicitly chooses to keep the currently saved value, reread it and use metadata-only `set_test_construction` with `status: supplied` and their retention statement. This records their choice to retain that value, not authorship of its wording; executable content stays unchanged. For cancellation alone without that choice, the current API has no clear-only operation. Do not write unsupported `status: saved`, falsely mark it missing, or invent agreement. Stop acting on the suggestion and explain that its persisted proposal marker remains; do not treat it as a renewed user request on a later read.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}