Example 1
Job: You have a transcript, notes, or a rough recording summary and need decisions, owners, deadlines, and unresolved questions without inventing promises.
Safe input boundary: Use a synthetic or consented transcript. Remove customer names, private compensation details, legal facts, and anything that should not leave the company.
- Ask AI to separate decisions, action items, open questions, and risks into labeled sections.
- Require every action item to include an owner, verb, due date or “date missing”, and quote/evidence from the notes.
- Review against the transcript before sending; do not let the tool email or schedule anything automatically.
Evidence link: The meeting-notes evidence log tracks setup friction, import limits, privacy/export notes, and action-item accuracy requirements before recommendations.
Open the related evidence log Example 2
Job: You have inbox snippets, voice notes, reminders, and half-finished ideas and need a realistic week plan plus reply drafts.
Safe input boundary: Paste a cleaned bundle rather than connecting Gmail, Outlook, Slack, Notion, CRM, or calendar accounts unless the account and automation scope are project-safe.
- Ask for exactly six sections: urgent replies, calendar/reminder plan, decisions needed, delegated tasks, content reuse ideas, and follow-up questions.
- Require the assistant to mark missing facts instead of pretending it attached files, updated checklists, sent messages, or changed calendar events.
- Copy useful drafts manually after checking dates, names, money amounts, and action claims.
Evidence link: The admin-routine evidence log includes scored prompt-only baselines for tools that produced full answers and setup-friction records for tools that gated or echoed prompts.
Open the related evidence log Example 3
Job: You have a messy client memo and need a clear deck outline, talking points, and exportable slides without losing the business logic.
Safe input boundary: Use a fictional or sanitized memo first. Avoid uploading confidential client data to a tool whose retention/export rules you have not checked.
- Ask for an 8-slide structure with one core message per slide, speaker notes, and a list of claims that need human proof.
- Check whether the tool can produce a private draft and export to PPTX/PDF before investing time in design polish.
- Treat sign-in gates, stalled generation, and missing exports as setup friction, not output quality.
Evidence link: The presentation-tools evidence log currently records no-login/signup/security/export friction, including a SlideGen no-login composer run where no generated deck appeared.
Open the related evidence log Example 4
Job: You want a chatbot or internal assistant to answer repetitive support questions from a known FAQ without making up policy details.
Safe input boundary: Start with a small public FAQ or synthetic knowledge base. Do not connect live helpdesk, customer data, or billing systems until fallback and deletion controls are verified.
- Upload or paste only the allowed knowledge base, then ask test questions with exact expected answers and “should escalate” cases.
- Check grounding, citation/source behavior, fallback wording, handoff controls, transcript export, and deletion/training settings.
- Publish or recommend only after private-bot transcripts and answer-key scoring exist.
Evidence link: The chatbot evidence log maps public-doc availability, signup blockers, upload/export controls, and no-score setup-friction evidence before quality rankings.
Open the related evidence log Example 5
Job: You need a first website draft for a local service business: homepage copy, services, local SEO basics, contact sections, and editability checks.
Safe input boundary: Use a fictional business brief first, and do not connect a real domain, payment plan, or customer analytics until export/lock-in and ownership rules are clear.
- Give the same bounded business brief to each builder so you can compare like with like.
- Check mobile layout, invented testimonials/certifications, local SEO basics, editing time, export/migration limits, and unpublish/delete controls.
- Keep draft-site screenshots and raw notes before assigning quality scores.
Evidence link: The website-builder evidence log uses one fictional cleaning-business brief and separates generated-site quality from signup, payment, domain, and export friction.
Open the related evidence log