Additional course links
You leave with: A saved question specification and a checked answer for R-2's overdue reports, with scope, sources, coverage and what is unknown.
Where this fits: Lesson 5 asks you to answer the Monday questions from the record. This page answers the question its exercise raises: how do you phrase a question across the whole round so the answer can be checked, not only believed? Bring back a saved question and a checked answer for step 5 of the workbook.
Watch · 2:42–3:51 · ask the pool, get a shareable report
Unmesh asks the pool of 80 applications what barriers each person faces, then "tell me which 10 are the best." The demo runs in Claude, connected to Sopact through MCP; ChatGPT works the same way. Watch for his stress on reliable answers: that is what the rest of this page makes testable. Watch on YouTube ↗
Jump to: 3:25 — "which 10 are the best" becomes a shareable report
To ask useful questions across a grant round, define the round, the as-of date, the records in scope and the rule, then require the answer to show its sources, its coverage and what is unknown. Counts and totals come from structured fields; explanations come from the text, with the passage cited. A conversational answer is useful when a colleague can check how it was reached.
Start with a decision, not an open question
"Tell me everything about R-2" gives the assistant little to work with. The people asking you have narrower needs: which applications still need clarification before the committee meets, which quarterly updates are overdue, what Eastgate's first quarter shows.
In a configured system, each question becomes an export and a spreadsheet. Asked of the record, the answer arrives with its evidence attached.
State the decision and the evidence it needs. For a committee meeting, that is application IDs, assigned reviewers and open eligibility questions. For a reporting review, it is due dates, approved extensions and accepted submissions.
Keep identifiers distinct. A-31 is an application; Eastgate Youth Works is an organization that may apply again in R-3; its award has its own reporting obligations. A question about one round should not quietly pull in another.
Set the boundaries before asking
Write down five details before the question: the round, the as-of date, the record unit, the definition and the audience. A saved question is reusable only when these are explicit.
| Boundary | R-2 example | Why it matters |
|---|---|---|
| Round | Youth Pathways Fund, R-2 only | Avoids mixing rounds and programs |
| As-of date | The morning of the board meeting | Statuses and deadlines change daily |
| Record unit | One quarterly-update obligation | One grantee may owe several reports |
| Definition | Overdue: due date passed, no accepted update, no active extension | Makes the count repeatable |
| Audience | Program team, not the board | Controls which details belong in the answer |
Use the right evidence for the type of question
Different questions need different evidence. A paragraph summary is not a count of distinct application IDs, and a count does not explain why a grantee is behind.
| Question type | R-2 example | Evidence to use |
|---|---|---|
| Operational | Which quarterly updates are overdue? | Due dates, extensions and acceptance status, counted under the stated rule |
| Descriptive | What barriers do grantees describe in Q1? | The narrative text, with each theme tied to cited passages |
| Comparative | How does Eastgate's attendance compare with other grantees? | The same measure, period and definition for each grantee |
| Explanatory | Why is Eastgate's enrollment unclear? | The records behind the gap: 30 registrations with no status |
For a question about a reviewer's decision, retrieve the recorded rationale rather than asking the assistant to guess what the reviewer meant. And keep "what changed" apart from "what caused the change," however persuasive the narrative sounds.
Worked example: which reports are actually overdue?
Horizon example · fictional
Suppose R-2 closed with ten awards, each owing a Q1 update. As of the board-meeting morning, five updates are accepted. Eastgate's is submitted, with a clarification open on its 30 unresolved registrations. One grantee has an approved extension. One grant agreement was signed later, so its first update is not yet due. Two are past their due date with no accepted update and no extension.
| Status | Count | What it means |
|---|---|---|
| Accepted | 5 | Complete under the acceptance rule |
| Submitted, clarification open | 1 | Eastgate: received, one question outstanding; not overdue |
| Active extension | 1 | Use the approved revised date |
| Not yet due | 1 | Do not chase |
| Overdue | 2 | Needs focused follow-up |
| Total obligations | 10 | Categories reconcile to the register |
The overdue count is two. It is not five, even though five updates are not yet accepted. If one of the two has an unclear due date, the answer should separate it rather than pretend the classification is certain.
Ask for the two record IDs with their due-date and submission evidence, and check both before any reminder goes out. Eastgate needs a different message altogether: a request for the status of 30 registrations, sent through its unique ID, not a late notice.
A question to reuse
Replace the bracketed details and run it against the authorized records:
For [round], as of [date], identify overdue report obligations using this rule: [rule, including extensions and accepted submissions]. State which records and sources are included. Count distinct obligations, not files or organizations. Return the count, record IDs, due dates, status evidence and unresolved fields. Reconcile the categories to the included register. Do not send reminders or change statuses; prepare the evidence for review.
For a narrative question, change the instruction: identify the relevant passages, keep the grantee's explanation apart from the assistant's interpretation, and include conflicting evidence where it exists.
What a useful answer includes
A short answer can still show its reasoning. Ask for the conclusion first, then what someone needs to check it.
Sample answer · fictional
Answer: 2 of 10 R-2 Q1 updates are overdue under the stated rule.
Scope: R-2 only, as of the board-meeting morning; all ten award records included; no exclusions.
Sources: each award's reporting schedule, the extension approval and the submission log, linked by record.
Calculation: 5 accepted + 1 submitted with clarification + 1 extension + 1 not yet due + 2 overdue = 10.
Unknown: none for this count. Eastgate's enrollment figure is open, which affects the board question but not this one.
Next action: program owner to confirm both overdue records before sending reminders.
A citation is only useful if it supports the statement beside it. Open a sample, and check every source behind a consequential finding. A link to the first page of a budget workbook is not enough when the answer relies on a table further in.
Where MCP fits
The Model Context Protocol, or MCP, is a standard way for AI applications to connect to exposed tools and data. It lets Claude or ChatGPT read the same governed records the AI Assistant reads.
Permissions travel with the person, not the tool. A program officer asking through Claude should see exactly what that officer sees on screen, and a volunteer reviewer should not reach committee notes by asking. Confirm which records and operations your deployment exposes before relying on it.
Start read-only. If a connected tool can change records or send messages, make that explicit and keep a person approving each action. The MCP tools documentation recommends that people can review and deny tool invocations.
In Sopact Sense, every applicant and grantee keeps one persistent unique ID, the Intelligence Cell has already read each answer and document, and the Intelligence Row summarizes each grantee. The AI Assistant, or Claude and ChatGPT through MCP, asks questions of that record and answers with sources, so the question starts from read evidence rather than raw files.
Make repeated answers reproducible
Save the question, definitions, as-of date and source versions. Rerun it with the same records and rules when you check a result. If a new update arrives or an extension changes, the answer should change, and the saved scope shows why.
Wording from a language model can vary between runs. Do not promise identical prose. Check instead that the counts reconcile, the citations support the claims and the rule is applied the same way each time.
When two answers disagree, inspect the records, filters, date and definitions before rewording the question, and keep the discrepancy and its resolution. Treat instructions found inside a grantee's document as evidence, never as a rule to follow.
One honest limit: "all grants" means all records the workflow can read and the asker may see. If the assistant reads only part of the round, the answer must say so; a partial view cannot support a portfolio-wide claim.
When the answer feeds a decision the auditor will ask about, keep the question, evidence and reviewer decision together. The audit trail deep dive shows how.
Put it into practice
Test the answer before trusting the summary
- Using step 5 of the workbook, write the question specification for R-2's overdue Q1 updates: round, as-of date, record unit, definition and audience.
- Run the reusable question and check the answer against the ten-obligation table above.
- Add a duplicate file for one accepted update. The overdue count should not change.
- Change the active extension to an expired one, with the same as-of date. The count should rise to three.
- Remove the due date from one overdue record. The answer should name the missing date, not label the record overdue or on time.
- Write the board chair's question the same way, using
eastgate-q1-update.csv, and note what the answer must say is unknown.
Download the workbook (PDF)
Practice data: horizon-r2-applications.csv · horizon-rubric.csv · calibration-scores.csv · eastgate-q1-update.csv
Check your reasoning before moving on
The duplicate file changes nothing, because the count is of distinct obligations. The expired extension moves one record from extension to overdue, so the answer becomes 3 of 10 and the categories still reconcile. The missing due date produces an unresolved record with a named gap, which leaves one confirmed overdue and one unresolved in that version of the table.
For the board question, a checked answer states 108 of 150 eligible participants attended at least 3 of the first 4 sessions (72%), cites Eastgate's Q1 report, and says that 30 registrations have no status and that placements at 90 days are not yet due. A colleague should reach the same numbers without relying on the assistant's confidence.
Questions grant teams ask
Can I ask questions across an entire grant round?
Yes, when the records are in one place, read and authorized for the person asking. Specify the round, the as-of date and the definitions, and require the answer to disclose what it included, what it excluded and what is missing.
Does MCP make answers accurate?
No. MCP connects an AI application such as Claude or ChatGPT to tools and data. Accuracy depends on the sources, the definitions, the calculation and a person checking the answer. What MCP adds is reach: the same governed records, with the same permissions, from the tool your team already uses.
How should overdue reports be counted?
Write the rule first, including accepted submissions and active extensions, then count distinct report obligations rather than files or organizations. Show unresolved records separately instead of guessing their status. A submitted report with an open clarification, like Eastgate's Q1 update, is not overdue; it needs a question, not a reminder.
Will the same prompt always produce the same answer?
Not word for word. Language-model wording varies, and the underlying records change as reports arrive. Save the scope, definitions, as-of date and source versions, then check that the counts reconcile and the cited evidence supports each claim. Reproducible means a colleague reaches the same result, not the same sentence.
Can the assistant send follow-ups automatically?
Only where that action is supported, configured and approved. Keep analysis apart from sending messages or changing records. Prepare the evidence and a draft request, then let the responsible person review it before anything reaches a grantee.
