Version. 0.1
Date. 2026-07-29
Status. Plain-language companion to loomworks-ask-your-engagement-investigation-v0_1. Same content, simple words. Where the two differ, the investigation governs.
Author. Claude.ai. Operator: Marvin Percival.
We want a partner to be able to ask their engagement a question in ordinary words — "what did the customer references say about churn, and what does the contract actually promise?" — and get a real answer.
The answer would come from everything the engagement holds. That means the documents that were uploaded, and also the record the team built as they worked: the notes from reference calls, the figures that were corrected, the concerns that were raised, who said what and when. Every part of the answer says where it came from. And if the partner wants to keep the answer, it goes into the record like any other contribution — permanently marked as machine-assisted.
Tools like Hebbia read documents and answer questions about them. They do that very well. But a lot of what a deal team learns is never a document. A reference call is a conversation. A partner's doubt is a remark. A corrected number is a change over time. Document readers cannot see any of that, because it was never on a page.
Loomworks holds exactly that material. So Loomworks could answer the one kind of question no document reader can: the question whose answer is half in the paperwork and half in what the team learned by talking to people. That is the question we would demonstrate.
We are not trying to match Hebbia. Honest limits, stated as promises:
Right now, asking the engagement a question does not work properly, for two separate reasons. So there is a strict order:
1. First, make answers truthful. During testing, we asked a tiny engagement — only six notes — a question, and it answered with an old value that had already been corrected, presented as if it were current. That is worse than saying nothing. Before anything else, we find out why that happened and fix it. There is a complication: some recent work on exactly this may already exist, but our records for June and July are incomplete, so step one includes checking those records. We do not build anything on top of answers that can be wrong.
2. Then, make the engagement searchable. Today, when asked a question, the system simply looks at the fifty most recent notes — whatever they are. The question doesn't actually search anything, and no search machinery has ever been built. This step builds it: a proper index over the notes and the text of uploaded documents, so a question finds the relevant material wherever it sits. This is the biggest piece of work.
3. Then, make it conversational. Once answers are truthful and search exists, letting the partner ask in plain words — and keep the answer in the record — is the smaller final step, because the conversation machinery and the way contributions enter the record already exist.
Separately, and at the same time: the work already underway to accept spreadsheets and slide decks continues on its own track. It makes the answers richer but doesn't block any of the three steps.
When all three land, our standing rule — never demonstrate question-answering to a prospect — lifts, and the sales conversation gains its second live moment: create the engagement from the firm's own playbook, then ask it a question whose answer quotes a reference call.
DUNIN7 — Done In Seven LLC — Miami, Florida Ask your engagement — the plain-language version — v0.1 — 2026-07-29