Ground a product explainer
Build a claim ledger
Before course create, list the claims the explainer may make. For each claim, record:
claim
source name and current location
source date or observed date
support strength: direct | qualified | absent
allowed wording
The ledger is working state, not course copy. Do not retain private source text beyond the user's request, and never record credentials or signed URLs.
Prefer sources in this order:
- current user-supplied product documentation, pricing, security, and release materials;
- current first-party product pages and demos;
- user-confirmed internal briefs;
- current third-party evidence for externally verifiable claims.
A source's marketing confidence is not evidence strength. Cross-check time-sensitive pricing, availability, integration, certification, and roadmap statements against a current first-party source. Use exact dates when “current” could become ambiguous.
Choose the grounding path
Strict document grounding
Use materials_only when the user supplies authoritative PDF, PPTX, DOCX, Markdown, or TXT files
and the course must stay within them. Set declared_sources to the exact retained count, upload all
documents, and wait for canonical processing before advancing; while any source is pending or
processing, run advance is a 200 no-op and allowed_actions omits continue.
When supported, a concise UTF-8 Markdown or TXT product brief assembled only from verified facts is the preferred source for a website explainer. It should contain the claim ledger's allowed wording, audience, verified workflow, proof, limitations, and CTA.
Supplied text or researched brief
When no document can be uploaded, use knowledge_source_mode=open with a compact constitution in
topic and the verified brief in content. State explicitly that product facts may not exceed the
brief. Because the mode can supplement general knowledge, perform a sentence-level claim audit at
both outline_ready and script_ready.
Do not convert uncertainty into a polished assertion. Use “designed to,” “supports,” or other qualified wording only when that qualification accurately matches the source.
Safe page content
The course may explain:
- the verified audience and problem;
- the product category and supported outcome;
- a documented process, architecture, or user journey;
- supported use cases and differentiators;
- current evidence and limitations;
- a supplied next step.
The course must not manufacture:
- interface screens, clicks, menu names, or workflow steps;
- customer logos, quotes, case studies, adoption, or revenue;
- performance, ROI, conversion, time-saved, or accuracy numbers;
- integrations, platforms, file types, languages, or regions;
- pricing, discounts, trial terms, availability, or roadmap;
- compliance, privacy, security, certification, or legal claims;
- competitor facts, rankings, or “best” claims without current evidence.
If a useful field is absent, omit it or name it as an open question for the product owner.
Image boundary
Use exact supplied screenshots as Pins when their fidelity matters. Use References only when a new editorial image may borrow subject, palette, or composition.
For pages without verified product UI, direct the image system toward editorial diagrams, metaphors, workflows, objects, environments, or typography. Include an explicit exclusion such as:
Do not depict an application screen, dashboard, browser chrome, controls, metrics, customer logos,
or interface text.
Inspect generated images for implied facts. A convincing invented dashboard is still a factual failure and must be regenerated.