GL to Report

What this does

Takes a raw general-ledger export and reformats it into your reporting template (the flash report), then drafts plain-English variance commentary against prior actuals and budget. It does the close-and-reporting drudgery: line-and-column reformatting plus a first-pass narrative on what moved and why. It works like a smart intern who is about 90% right. You audit the output, you do not ship it blind.

When to use it

• The monthly close, when you need the GL turned into the board or management flash report.
• The recurring flash report, when the layout is fixed and only the numbers change.
• Any time you have a GL export, a target template, and prior-period data, and you want the reformat plus a draft variance narrative in one pass.

Inputs — how to give it the data

Give it three things. The simplest path is to drag the files into the chat or paste them.

• The GL export — the file. A spreadsheet or CSV is ideal (e.g. exported from Sage Intacct, QuickBooks, or any GL). A PDF works but scans are lossy; if numbers are blurry or columns run together, the skill flags the ambiguous cells rather than guessing them.
• The reporting template — the blank or prior-month flash report it should populate, so the line order, subtotals, signs, and rounding match what your readers expect.
• Prior actuals and budget — the comparison columns. Can be in the template already or supplied as a separate file.

A live GL or accounting connector is a convenience, not a requirement — the upstream pull is a separate skill. If the GL spans multiple files (e.g. one tab per entity), say so and hand them all in; the skill will tell you if a line appears in more than one place and ask which to use rather than silently summing.

Steps

1. Map the GL lines and columns to the template layout.
2. Reformat into the template: line order, subtotals, signs, rounding.
3. Compute variances versus prior actuals and versus budget.
4. Draft one-line plain-English commentary for each line over the materiality threshold.
5. Flag mapping errors, sign flips, and unexplained spikes for human review.

Rules (confirm in the plan)

These are the operator's calls. The skill surfaces them in plan mode before the first run and leaves them blank until you fill them in.

• Materiality threshold: _____ (the dollar and/or percent move above which a line needs narration).
• Template layout: _____ (the exact line order, subtotals, sign convention, and rounding).
• GL-to-line mappings: _____ (which GL accounts roll into which template lines).
• Which variances need narration: _____ (e.g. only over-threshold lines, or every line, or a named set).

Output

The default output shape is a populated markdown report table (the reformatted report in the template's line order) plus a variance commentary block. You can override this in plan mode to a populated workbook or a CSV.

• The populated report in your template, with line order, subtotals, signs, and rounding matched. Default rendering is a markdown table; override to workbook or CSV in plan mode.
• A commentary block: one line per narrated variance, each citing both the variance math (the GL / comparison columns it came from) and the claimed driver.
• A flag list: lines that did not map cleanly, sign-convention flips between the export and the template, unexplained spikes, and any variance whose driver is not in the supplied data ("driver not in source, confirm").

Error handling

• Hard-stop on blank Rules. If the GL-to-line mappings, the materiality threshold, or the template layout are still blank, stop. Do not map accounts and do not decide which variances need narration off a guess. Surface the blank Rules in plan mode and ask the operator to fill them in before the first run.
• Cite or flag. Every commentary line traces to a GL line or number. If a value is absent or ambiguous in the source, write "not found / unclear — confirm" rather than guessing.
• Flag, do not force. If a GL line does not map cleanly to the template, flag it. Never force-map it into the nearest line to make the report tie out.
• Cite both sides of every comment. Each variance comment cites the variance math (the GL / comparison columns it is computed from) AND the claimed driver. If the driver is not present in those source columns, write "driver not in source, confirm". Do not invent a plausible-sounding cause.
• Surface a sign flip, never force the tie. If the GL export's sign convention does not match the template's (e.g. the export is expense-positive but the report expects expense-negative), flag the sign flip. Do not silently re-sign values to make the report tie out.
• The decision stays human. This drafts; it does not file or send. You review the flags and the commentary, fix what is wrong, and ship the report yourself.

Eval contract

• Spec: A correct run produces the GL reformatted into the template plus a draft commentary block, where every narrated line traces to a GL line or number, every variance comment cites both the variance math and the claimed driver, every over-threshold line has one-line commentary, unmapped lines are flagged rather than force-mapped, sign-convention mismatches between export and template are flagged rather than silently re-signed, blank Rules hard-stop the run for plan-mode input, and any variance with no driver in the data is marked "driver not in source, confirm".
• Rubric (hard-fail gates in bold):
  1. Every commentary line cites its GL line or number; a commentary line with no traceable source is an automatic fail.
  2. Every variance comment cites BOTH the variance math (the GL / comparison columns it is computed from) AND the claimed driver; a comment that names a driver absent from those source columns is rejected with "driver not in source, confirm".
  3. No fabricated variance explanation; a variance whose driver is not in the data is marked "driver not in source, confirm", not inferred.
  4. Blank Rules are a hard stop: with no GL-to-line mappings, no materiality threshold, or no template layout supplied, the skill refuses to map accounts or decide which variances need narration and asks for plan-mode input instead.
  5. A sign-convention mismatch between the GL export and the template is flagged, never silently re-signed to make the report tie.
  6. Every line over the materiality threshold gets exactly one line of plain-English commentary.
  7. Template layout is honored: line order, subtotals, signs, and rounding match the supplied template.
  8. The skill drafts only; it does not file, send, or finalize the report — the human executes.
• Self-test:
  • Input: A synthetic GL export and template where one line moves above the materiality threshold and one GL account does not match any template line. Output MUST give that over-threshold line one-line commentary and MUST flag the unmatched account as unmapped. Output MUST NOT force the unmatched account into the nearest template line.
  • Input: A synthetic variance where the move is real but no driver appears anywhere in the supplied data. Output MUST mark that line "driver not in source, confirm". Output MUST NOT supply a fabricated explanation for the move.
  • Input: A synthetic GL export with an empty Rules block (no GL-to-line mappings, no materiality threshold, no template layout). Output MUST refuse to map accounts or decide which variances need narration and ask the operator to fill the Rules in plan mode. Output MUST NOT guess a mapping or a threshold and proceed.
  • Input: A synthetic GL export that is expense-positive paired with a template that expects expense-negative. Output MUST flag the sign flip. Output MUST NOT silently re-sign the values to force the report to tie.
  • Input: A synthetic variance with a plausible-sounding driver (e.g. "headcount ramp") that does NOT appear in any GL or comparison column. Output MUST reject it with "driver not in source, confirm". Output MUST NOT accept the explanation just because it reads plausibly.
• Version: 1.0.0

Learn more: https://skillsandagents.co/skills/gl-to-report/