What it does
Looks for the shapes a reviewer looks for in a journal, and returns a short ranked list of entries worth opening. Each one comes with its usual harmless explanation and the step that settles it.
A bundled script, je_screen.py, computes every test.
15 tests, in four groups
Worked out
These can pass.
- Every entry balances
- No entry is booked twice
- Nothing was recorded after the close
Timing patterns
- Backdated entries
- Period-end clustering
- Weekend and night
Amount patterns
- Round amounts
- Just under an approval limit
- First digits, by Benford’s lawNeeds 500 or more entries, and never flags a single entry.
People, accounts, descriptions and reversals
- Who keyed the manual entries
- Rare account pairings
- Manual entries on control accounts
- Manual credits to income
- Weak descriptions
- Reversals and reversal chains
How it stays fair
A flagged entry matches a pattern. It is not a finding that anything is wrong, and the report never words it as one.
A test that lacks its column is listed as not run, with the reason, never as a pass.
The report
Every run writes the same sections, in this order.
- A header with the entry counts and the tests run
- How the file was read
- The worked out tests
- The patterns
- The top 25 entries worth opening
- The rest, summarised by test
- What was not run, and why
- A suggested order
What it reads
One row per journal line, with at least a date, an account and an amount, as CSV or Excel.
More columns turn on more tests: entry number, source or type, the date recorded, posted by, approved by, description and counterparty.
It also reads ledger-card exports, one sheet or block per account, and flattens them first. A trial balance cannot be screened: the script refuses it with a clear message.
Try “Screen my June journal entries and show me the ones worth opening. The period was closed on 5 July.”
Connected to Cybooks
Reads the journal entries and lines, the manual journals and the period’s close date from Cybooks, then runs the same script.
The people tests are reported as not run, because the journal read carries no posted by.
What you need
- Claude, with skills
- Anthropic offers skills on its Free, Pro, Max, Ultra, Team and Enterprise plans. To add this one, open Customize › Skills, press +, then Create skill › Upload a skill, and pick its file.
- Code execution
- Turn on Code execution and file creation in Settings › Capabilities, or use Claude Code. Every figure is computed by code, not estimated by the model.
- On its own
- CSV or Excel exports. It sees only the files you attach.
- Connected
- The Cybooks connector at https://app.cybooks.com.cy/mcp/cybooks, signed in. It reads with your own permissions: view reports, view the books review and view the dashboard. Two optional steps also need manage the period lock and view manual journals. The connector is on Premium.
How it was tested
Tested with 31 script tests, a real six-month general ledger export, and two rounds of end-to-end runs by a fresh Claude agent with only the skill file and the raw export. The changes from the second round and from a code review (13 issues) are covered by the script tests, not yet by a fresh end-to-end run. Not yet tested: a run inside the Claude apps, and the connected path against the live Cybooks connector.
Claude is a trademark of Anthropic. Cybooks is not affiliated with Anthropic; no endorsement implied.