Guide the Senso v2 onboarding experiment: understand the brand, bring in relevant knowledge, compare a recorded external AI answer with approved facts, stage a checked draft with evidence, then publish when requested. Use for Senso quickstart v2, testing the new verification loop, or continuing a v2 onboarding session. Ordinary file saves belong to senso-context-layer.
npx @senso-ai/shipables install senso-ai/senso-quickstart-v2Requirements: Requires Senso CLI 0.17.0 or newer, authenticated organization access, and the senso-context-layer, senso-evaluate-remediate, senso-generate-verify and senso-publish skills. Industry tracking requires GEO access.
This is a beta orchestration skill. It guides one verification loop, not an open-ended campaign. Lead with the useful outcome: understand one AI answer about the user's industry, compare it with approved knowledge, and prepare a draft with evidence. Publication and later citation are separate outcomes; neither is guaranteed by ingestion or a passing evaluation.
Use the four module skills for command mechanics. Load each before its stage. If a module is absent, explain which capability is unavailable and help install the official module when authorized. Do not silently fall back to the original quickstart, whose decisions differ from this experiment.
Within this v2 flow, these instructions govern sequence, question selection, review and reporting; module instructions govern commands, payloads and polling. Do not let the old module handoff restart a different onboarding flow. Respect a user's narrower request and existing authorization.
Run senso whoami --output json --quiet. Require a successful live identity, not cached information
returned after a network failure. Show the organization name. If the user named another organization,
stop organization-dependent work until credentials are corrected. Never request a key in chat; direct
the user to senso login locally. Do not print credentials.
Read organization settings and relevant existing state using the modules. Resume the selected question and its draft rather than guessing the stage from one sample item. Keep identifiers across the loop: org, question, answer/run, source, content/version, evaluation, and publication. Do not reset settings.
For a new brand, say:
You're connected to [organization]. We'll bring in your knowledge, compare one AI answer with your approved facts, and prepare a draft with evidence. What's your website?
Reuse a previously confirmed website. If there is no site, accept a user-named brand document or a brief description of what the brand does, who it serves and what it wants to be found for. Ask only for information still needed. Do not scan personal folders for material.
Use senso-context-layer and senso-evaluate-remediate.
If there's already a useful source article, start from it and map it to a relevant tracked question. No repeated full-site imports merely to run the demonstration.
Exit: brand understood, relevant question selected, supporting knowledge retrieved. Preserve the question ID. If GEO is unavailable, demonstrate source-backed retrieval and clearly stop the external-tracking path rather than claim a complete four-step run.
Read External answers for the verified CLI retrieval path. Capture the question, model, answer date, exact excerpt and cited URLs when available. A supplied answer is user-supplied evidence, not a new live run. Never manufacture an external answer.
Compare atomic claims with approved source passages. Show the most useful finding in plain language: The answer says … The source says … The issue is … Include the source reference.
Distinguish:
If knowledge is missing, ask for the smallest needed fact and evidence. Through the context module, save authorized source contributions, confirm processing and retrieve them before using them in a draft. If approved sources conflict, surface that conflict; don't pick whichever supports the demo. For comparison probes, suppress search gap signals where supported; record a real issue deliberately.
Exit: one evidenced issue or an explicit explanation of unavailable evidence. Never substitute an internal KB answer for the external-model comparison without telling the user.
Use senso-generate-verify. Read brand guidance, content type and source passages. Write a concise
draft addressing the selected issue. Facts newly supplied by the user go through the knowledge
contribution path first. Do not add unsupported material to improve the score or story.
Show the draft and evidence for review. If the user already asked to create staged drafts, save the private draft without another save-permission question. Otherwise follow the module's review-before-save flow. Store publishing content in Drafts, not as another KB source. Confirm content ID, version and status by readback. A new version is not a new distinct draft.
This v2 verification workflow includes both factual and brand checks. Run the supported kb_accuracy
and brand_alignment evaluators when a meaningful brand kit exists. If alignment cannot be checked,
state that it is unassessed; an empty-kit pass is not evidence. Read the actual result, claim coverage,
source evidence and failures. Process completion or a zero exit code is not a passing quality result.
Present a concise receipt using Receipt and checks. Link checks to the saved version. If native receipt export is unavailable, label the artifact a compiled verification summary. Do not claim a signed or native receipt exists. Do not equate KB grounding with legal compliance.
Fix failed claims with evidence or remove them in the review. Do not dismiss a separate knowledge gap just because a sentence was removed. Re-evaluate changed content; after a second unresolved evaluation, leave the draft staged and show the blocker rather than loop indefinitely.
Exit: a visible staged draft, sources, version-linked checks and explicit unresolved items. An internal correction does not mean ChatGPT, Gemini or another public model has changed its answer.
Use senso-publish. Confirm exact draft/version, verification status, destination, and CTA. Check that
public content and evidence do not disclose internal material. Draft approval alone is not permission
to publish. If the user already authorized this version and destination, don't ask again.
Publish the evaluated version using documented mechanics. Read each destination's returned status and verify the live page. Report partial failures explicitly. Show the page and supported verification record reference. Do not invent a receipt URL.
Record a dated baseline for the same question, models and market. Distinguish mention rate, owned citation rate, referrals and completed transactions; use actual denominators. No observations means unavailable, not zero. A verification receipt is not a transaction receipt without real event linkage.
Close with what exists and what remains observable:
The checked page is live at [URL]. [Verification reference.] We haven't yet observed a new citation. Next we can compare the same question over time, or work on another answer gap.
Do not promise a citation timeframe or improved rankings. Don't create a recurring monitoring job unless requested. If publication wasn't authorized, close on the saved Drafts entry and its review status.