Quarterly Commentary Prompt Chain: Four Prompts From Your Figures Table to a Checked Draft
For Investment Fund Managers ·
What This Builds
A quarterly letter goes wrong in two places: a figure that does not match the performance system, and a sentence that says more than the figures support. This chain is four prompts run in one Claude chat, with a checkpoint after each. The first restates your figures table and lists what the commentary template asks for that the table does not hold. The second drafts three sections from the table and your notes only. The third matches every number in a draft to the table. The fourth lists every phrase that reads as a comparison, a claim, a forecast or a recommendation, so your CCO sees them first. The outcome it serves is a draft in which every figure traces to your system and every claim has been shown to the person who decides on it.
Prerequisites
- A Pro subscription ($20/month), or another plan your firm has approved for this work. That is the total ongoing cost.
- Your quarter's figures table, copied from your performance system, and your own positioning notes. While you learn the chain, use the invented table in this guide.
- Your commentary template, or a list of the figures and sections it asks for
- A word with your CCO on which AI tools are approved and how chats are kept
The Concept
A good editor does not rewrite a draft in one pass. The editor reads the figures first, then the prose, then asks what each sentence claims. This chain splits the work the same way. Each prompt does one small job and stops, and you check the result against the source before the next one starts.
One rule runs through all four prompts. The model never calculates a number, never adds one, and never supplies a reason you did not give it. Where the table or your notes are silent, the model writes a marked gap and you fill it from your system. The spreadsheet and the performance system do the arithmetic, and your CCO decides how returns are presented and which words are acceptable.
Data caution. Material non-public information never goes into any AI tool. Holdings that are not yet disclosed, trade ideas and orders, your models, investor and client names, account numbers, fee terms and side letters stay in your firm's systems unless the firm approved this tool for that data and your CCO allows it. Figures that come from a data provider or a licensed research service carry licence terms, so read them before you upload any of that content. The chain works on invented data, on figures your firm already uses in approved text, and on column names alone. Before real figures go in, check the tool's data settings for whether chats are used for training and how long they are kept, because a free consumer plan may use chats to train models unless the setting is off. Ask your CCO which tools are approved and how business chats are kept.
Build It Step by Step
Part 1: Prepare your inputs
Open a new Claude chat for the quarter. Keep one chat for the whole chain so each prompt can see the earlier answers.
Use the invented Strategy S figures below for your first run. Every number is invented, and the contributions are in percentage points.
| Item | Invented value |
|---|---|
| Net return, Strategy S (percent) | 1.8 |
| Gross return, Strategy S (percent) | 2.1 |
| Return, Benchmark B (percent) | 1.5 |
| Difference, net minus Benchmark B (percentage points) | 0.3 |
| Top contributor, Company C (percentage points) | 0.6 |
| Top contributor, Company F (percentage points) | 0.4 |
| Top detractor, Company A (percentage points) | -0.5 |
| Top detractor, Company H (percentage points) | -0.2 |
Invented positioning notes: "Added to Company D. Trimmed Company C after its run. Sold Company J after the thesis broke (management change)."
Invented draft paragraph from a colleague, with two planted figure errors, for Prompts 3 and 4: "Strategy S returned 1.9 in the third quarter, ahead of Benchmark B at 1.5. Company C and Company F led, with Company F adding 0.5. We expect the portfolio to keep outperforming as our holdings deliver."
Invented list of what the commentary template asks for: net return, gross return, Benchmark B return, difference net minus Benchmark B, top two contributors, top two detractors, year-to-date net return, since-inception net return, number of holdings, cash weight.
Part 2: Prompt 1, restate the table and list the gaps
Paste the table, then the template list, into the prompt where the brackets are. Send it.
Below are two things: a figures table and a list of figures a commentary template asks for.
First, restate the figures table exactly as a table. Keep every value, label and unit as written. Do not round, convert, reorder or add anything.
Second, list every figure on the template list that the table does not hold, writing "not in table" next to each one. Do not estimate any of them and do not compute any of them.
Figures table:
[paste the figures table]
Template list:
[paste the list of figures the template asks for]
Checkpoint 1. Compare the restated table with your source, line by line, value by value. With the invented data, the output should repeat all eight values and mark four items "not in table": year-to-date net return, since-inception net return, number of holdings and cash weight. If any value changed, stop and rerun the prompt. A restated table you have not checked is not a source.
Part 3: Prompt 2, draft the sections from the table and the notes
Paste your positioning notes where the bracket is, and send.
Using only the figures table above and the positioning notes below, draft three short sections: "The quarter in brief", "Contributors and detractors" and "Positioning changes".
Rules:
- Use every figure exactly as it appears in the table. Do not compute, round or add a figure.
- State values plainly. Do not use words that compare, rank or judge, such as ahead, behind, above, below, led, lagged, strong or weak.
- Describe each positioning change only as the notes describe it. Where the notes give no reason for a change, write "[reason: PM to supply]".
- Where a figure is needed and the table does not hold it, write "[figure from the performance system]".
- Write no forecast, no outlook and no recommendation.
- Label the output "Draft for PM and CCO review".
Positioning notes:
[paste your positioning notes]
Expect the quarter-in-brief section to state the net return, gross return and Benchmark B return side by side, the contributors section to name Company C and Company F with their values and Company A and Company H with theirs, and the positioning section to describe the three moves from the notes, with "[reason: PM to supply]" for the Company D addition. Your wording will differ.
Checkpoint 2. Read the draft against the table and the notes. Every figure must appear in the table. Every move must appear in the notes. Delete any sentence that adds a reason, a view or a figure you did not give. Fill each marked gap yourself, from your system or your own judgment.
Part 4: Prompt 3, match every number to the table
Run this on the draft you want checked. In the first run, paste the colleague's paragraph from Part 1. In your real work, paste your own draft, including the one you just edited.
List every number in the draft below. For each one, give the matching entry in the figures table above and write Match yes or no. If the number has no matching entry, write "no entry". Report only. Do not correct the draft, do not compute anything and do not explain.
Output a table with three columns: Number in draft, Table entry, Match.
Draft:
[paste the draft]
On the colleague's paragraph, the correct output has three rows. The number 1.9 is a no, because the table's net return is 1.8 and the table holds no 1.9. The number 1.5 is a yes, matching the Benchmark B return. The number 0.5 is a no, because the table shows Company F at 0.4. On the draft from Prompt 2 and Checkpoint 2, every row should read yes.
Checkpoint 3. Do not take the Match column on trust. Open your performance system or the source table and confirm each row yourself. Fix every no from the source, never from the model. The model reports and you correct.
Part 5: Prompt 4, list the phrases for your CCO
List every phrase in the draft below that reads as a comparison, a claim, a forecast or a recommendation. Quote each phrase exactly and name which of the four it is. Do not rewrite the draft, do not suggest alternatives and do not say whether a phrase is acceptable.
Draft:
[paste the draft]
On the colleague's paragraph, the list should include "ahead of Benchmark B" as a comparison and "keep outperforming" as a forward-looking claim. It may also list the clause "as our holdings deliver" or the word "led". Extra items are fine, since your CCO decides what stays. A missing item is the problem, so read the draft yourself for any sentence the list skipped.
Checkpoint 4. Send the draft and the list to your CCO. Your CCO decides how returns are presented and which wording is acceptable. Each phrase on the list is a question for that person and not a finding of the model.
Part 6: Test the whole chain against something outside Claude
Run the chain once on a past quarter whose approved letter you already have. Compare three things with that letter: the figures Prompt 3 marks yes against the figures in the approved letter, the phrases Prompt 4 lists against the wording your CCO changed or removed that quarter, and the gaps Prompt 1 marks against what you had to fetch from your system. If the chain misses a phrase your CCO flagged, add a line to Prompt 4 that names the pattern in your own words and rerun.
Real Example: Strategy S, Q3 2026 (all data invented)
Setup: The invented table, notes, template list and colleague paragraph from Part 1. Input: Four prompts in one chat, with a check after each. Output:
- Prompt 1 restates eight values and marks four items "not in table".
- Prompt 2 returns three sections labelled "Draft for PM and CCO review", with "[reason: PM to supply]" at the Company D addition and "[figure from the performance system]" wherever the template needs a year-to-date or since-inception figure.
- Prompt 3 on the colleague paragraph returns three rows: 1.9 no (the table says 1.8), 1.5 yes, 0.5 no (the table says 0.4 for Company F).
- Prompt 4 on the same paragraph lists "ahead of Benchmark B" and "keep outperforming" for the CCO.
Check it by hand. The paragraph holds three numbers: 1.9, 1.5 and 0.5. The table's net return is 1.8, so 1.9 has no match. The Benchmark B return is 1.5, so that one matches. Company F's contribution is 0.4, so 0.5 has no match. The difference of 0.3 is the table's own entry, net 1.8 minus Benchmark B 1.5, and the model did not calculate it. What you do next: You correct the two figures from the performance system and fill the gaps. Then you send the draft and the phrase list to your CCO. The draft goes out only after your firm's marketing review approves it. The data is invented. Your real table comes from your own systems.
What to Do When It Breaks
- The restated table differs from your source → Rerun Prompt 1 with the table pasted as plain text. Check for merged cells or a pasted chart. If a value still changes, use the source table and skip the restatement.
- A draft contains a figure that is not in the table → The model added it. Delete it, then rerun Prompt 2 with the rule "use only values in the table" placed first. Prompt 3 will show it as "no entry".
- Prompt 3 says yes for a number that is wrong → The check is only as good as the table you pasted. Compare the table with the performance system, and check each Match yes by hand.
- Prompt 4 returns an empty list on a draft you know has claims → Read the draft yourself and send the sentences you doubt to your CCO. An empty list from a model is not clearance.
- The chat drifts or repeats earlier text → Start a new chat, paste the table and notes again, and run the chain from Prompt 1.
- You skip a checkpoint when pressed for time → The chain only adds safety if you check each output. Treat a skipped checkpoint as an unchecked draft, and tell your CCO.
Variations
- Simpler version: Run Prompts 3 and 4 only, on a draft you wrote yourself.
- Extended version: Raise the aim from a checked draft to a checked pattern. After each quarter, keep a list of every figure error and every phrase your CCO changed, and add the recurring patterns to Prompt 4 in your own words. Compare the next quarter's list against the last.
What to Do Next
- This week: Run the chain on the invented data, then on a past quarter with an approved letter.
- This month: Run it on the current quarter's draft, with the CCO review built in.
- Advanced: Move the prompts into a Claude Project with your template as project knowledge, and keep the four checkpoints.
Advanced guide for investment fund manager professionals. These techniques use paid features and may require a plan your firm has approved. Every output is a draft for the portfolio manager, and anything an investor, consultant or the public will see goes through your CCO's marketing review first. Claude updates its screens often, so if a label differs from this guide, look for the closest match.