factual-accuracylisted
Install: claude install-skill alectivism/organization-ai-skills
# Factual accuracy
Every specific claim in a deliverable traces to a source read in this session: a
file, a tool result, a page actually fetched, a transcript, or the user's own
words in this conversation. Recall is not a source. Being fairly sure is the
exact state in which errors get written confidently.
These documents go to customers, members, partners, or prospects deciding
whether to trust your organization. One invented attendance figure turns the
reader's question from "should I trust this" to "what else in here is wrong."
A gap is fixable. A confident invention is a liability nobody downstream can see.
## Sort every claim before writing it
| State | Test | Action |
|---|---|---|
| **Verified** | Traced to a source opened this session | Write it |
| **Inferred** | Reasoned from verified facts | Write it only with the inference visible in the prose ("their Q3 filing implies"), never as a bare fact |
| **Unknown** | Cannot be traced | Do not write it. Collect it for the gap ask below |
## Check your organization's own sources, never memory
These change, and asserting them from memory is the likeliest way to be wrong in
a way a reader notices.
| Claim | Source |
|---|---|
| Membership/customer status, join and renewal dates | Your own membership or CRM skill, if you have one |
| What a program, tier, or offering includes, what it costs, approved value wording | Your own catalog/offerings skill, if you have one, and any proof-points reference inside it for any numb