Skip to content

Month-end close pack: reconcile and summarize with Claude

Paste this month's transactions and your bank statement in, and get a numbered list of exactly what to check before you call it closed.

One email, and it's yours

The download starts as soon as you send it.

No download after a few seconds? Write to hey@vetted.tools and we'll send it.

For the Friday brief, a weekly email from vetted. Unsubscribe anytime. Privacy

What’s inside

  • reconcile-and-flag

    Checks a transaction export against a bank statement and lists duplicates, mismatched amounts, category gaps and missing receipts.

  • close-summary

    Drafts a plain-language note on what moved this month and a close checklist to hand to your bookkeeper.

How we tested this

Across 36 answers (12 made-up jobs, 3 tries each), the skills caught 65 of 66 planted mistakes (98%) and passed 123 of 124 checks (99%). The 22 mistakes hidden in the jobs count once in each try. Our bar, set on September 30, 2026, before any answer existed, was 80% on both. It cleared both. Tested October 20, 2026.

Tested

Planted mistakes caught

Above our 80% bar

65of 6698%

22 hidden in the 12 made-up jobs, counted once for each of the 3 tries.

Checks passed

Above our 80% bar

123of 12499%

Things every answer must also get right, or leave alone, like no made-up prices or dates: 41 in the answer key, counted once for each try, plus 1 for answers that broke a hard rule.

What a planted mistake is

Something we hid in a made-up job on purpose, a problem or a detail that's easy to miss, that the skill's own instructions say it must notice and deal with. For example: a landscaper's books show the same $412.00 Valley Turf Supply charge on the 5th and again on the 6th, but the bank shows it once. A right answer flags it as a possible double entry and names both dates.

What went in: made-up job 1

Landscaper, ledger duplicate

WAVE EXPORT

…

planted mistake: 09/05/2026,Valley Turf Supply,412.00,Materials,receipt attached

planted mistake: 09/06/2026,Valley Turf Supply,412.00,Materials,receipt attached

…

BANK STATEMENT (First Prairie Bank checking)

…

09/05/2026,VALLEY TURF SUPPLY,-412.00

…

What came out: the pack’s first try

#
1
Type
where the answer caught it: Duplicate
Line(s)
where the answer caught it: 9/5 Valley Turf Supply $412.00; 9/6 Valley Turf Supply $412.00 (both in Wave)
What to check
Same vendor, same amount, one day apart. … Your bank statement shows only one $412.00 Valley Turf charge (9/5), so the 9/6 entry may be a double entry to delete. …

…

2 items to check before this month closes.

This answer’s grade: it caught 2 of 2 planted mistakes and passed 4 of 4 checks.

The planted mistakeWhere the answer caught itBoth quoted from the test’s own files. … marks a cut.

Where it fell short

1 planted mistake missed1 check failed

Where it still slips, and one limit on purpose:

  • In 1 of 3 tries at one job, it said the reason for a jump in spending was unclear, then guessed at one anyway.
  • In 1 of 3 tries at another job, it used the other skill's name, "reconcile-and-flag", without saying what reconcile means. The other 2 tries did the same and passed, so the grader isn't perfectly consistent.
  • On purpose, the summary skill won't add lines up or work out a percentage, even when you ask. Your bookkeeper does that.

Try it on your own jobs first

The test used made-up jobs, not yours. In your first week:

  • Run the skills on a few jobs you've already handled.
  • Compare their drafts with what you actually sent.
  • Read every draft before it goes to a customer: you decide what goes out.

How the test ran

  1. Claude Opus writes the test

    Opus 5.5, in a conversation of its own, wrote 12 made-up jobs and an answer key, with 22 mistakes hidden on purpose. Its folder held only the two skill files and a written brief, and it could not open files outside that folder.

    Claude Opus 5.5

  2. Claude Sonnet does each job 3 times, using the pack

    Sonnet 5.5, a different version, ran the skills on every made-up job 3 times, each in a fresh conversation: 12 jobs × 3 tries = 36 answers. It never saw the answer key.

    Claude Sonnet 5.5

  3. Claude Opus grades every answer

    Opus 5.5 again, the model that wrote the test, in a new conversation, checked all 36 answers against the answer key, without being told which try it was reading.

    Claude Opus 5.5

Read the details
What it does not prove

A pass says the skills follow their own instructions on jobs like these. It does not say they will be right about your business.

  • The jobs are made up, and 36 answers from one model on one day is a small test.
  • We tried to leave out jobs where the skills' own instructions don't clearly say what to do, so this test can't tell you how they handle those. The list is in METHOD.md, in the download.
  • The grader is also an AI, the same model that wrote the jobs and the answer key, so the two can share blind spots, and no person checked its grades.
  • Every answer and every grade is in the download, so you can check them yourself.
What a check is

Something a right answer must also get right, or must leave alone: no made-up prices, dates or promises, no false alarms, the format the skill promises. The answer key has 41, counted once in each of the 3 tries. Some lines in the skill are hard rules, things it must always or never do. An answer that broke one got an extra check, marked failed, so a broken hard rule costs more. That happened once.

Where the made-up jobs came from

Claude Opus 5.5 wrote the 12 made-up jobs and their answer key, before any answer existed. Its folder held only the two skill files and a written brief, and it could not open files outside that folder. The businesses, people and numbers are invented. None of it is client data.

Who answered, and who graded

Two versions of Claude, both made by Anthropic, took part. Claude Sonnet 5.5 did the jobs using the pack, 3 times each, because Claude can answer the same job differently each time. Every try was a fresh conversation holding only the skill's instructions and the job. It never saw the answer key. Claude Opus 5.5, the version that wrote the jobs and the answer key, then marked every answer in a separate conversation, without being told which try it was reading.

What the test cost us

Running the jobs and marking the answers cost us about $2.08 in AI fees, at Anthropic's list prices in US dollars. That leaves out writing the test. That cost was ours. The pack is free.

The full detail

The download holds every made-up job, every answer, every grade and the full method, in its eval folder (start with METHOD.md).

Want this set up for your business?

Start with a free 15-minute call.

Talk to usPick a 15-minute slot(opens our calendar in a new tab)