DEV Community: JH-A

AI Office Workflows That Actually Work: Verified Guides for Excel, PDFs, and Documents (2026)

JH-A — Sun, 15 Nov 2026 14:10:18 +0000

A guide is only worth your ten minutes if it can tell you the file it used and the number it got. Every guide in this room starts that way: one file I opened, one job I tried to finish, one measurement I could not talk my way around, and the change that finally made the number right. Nine of them are published as of November 15, 2026, and this page is the map of all nine.

The work behind them is ordinary office work — an export someone emails you, an AI answer that has to become a Word document, a recurring call between two cities, a briefing you read before a meeting. I am an architect, not a developer, so the tests are sized for a person at a desk with a spreadsheet open and an AI chat in the next window: one file, one afternoon, one number to check at the end.

Nothing on this page ranks AI tools against each other. I do not run model benchmarks, and I would be guessing if I told you which assistant is "best at Excel". What each guide settles is narrower and more useful: given this file and this task, what does a plausible-looking answer get wrong, and what proves it wrong.

Key Takeaways A cleanup that ignored currency symbols reported 21,748,113.34 for a file whose real USD total is 51,464.59. The checksum caught it; the AI did not. The same AI answer pasted four ways gave Google Docs a 9×4 table with 36 of 36 cells and gave a Word file heading styles with zero cell borders. The clipboard decided that, not the model. Nine guides share one rule: the number was measured on a file, the date is recorded, and the guide states what it did not test. The 60-row CSV from the spreadsheet guide is public. You can run your own AI on it and compare against a published answer key before you trust that AI with a real export.

What does "verified" mean on this site?

Verified means three things are on the page: a file that exists, a number you can reproduce, and a sentence about what the test did not cover.

The first is the easiest to fake and the easiest to check. Where a file was synthetic, the guide says so and, in the spreadsheet case, hands you the same CSV to download. Where a file was someone's real export, the guide describes its shape — four date formats, three currency notations, 60 rows — without reproducing anything that identifies the sender.

The second is the number. If a post says an AI overstated a total by 21,696,648.75, that figure should be reachable from the data in a couple of minutes: sum the dollar column, sum the won column, compare against what the tool reported. A claim that cannot be rebuilt from the file is a claim I try not to publish.

The third is the limit, and it is the part most AI writing skips. The design review is a constructed before-and-after, not a study of which tool designs better. The keyword test used nine invented headlines, not a live feed. The gamification template borrows a mechanism from a running game; nobody has measured whether renaming a work task makes you faster, and the guide says so in its own key takeaways. My editorial policy is the long version of this paragraph.

That standard also decides what does not appear here. A topic I could not reproduce with a number stays unpublished, which is why the PDF section below is thin.

Which guides cover spreadsheets and messy data?

The spreadsheet guide is the one to start with if a file lands in your inbox that was exported by someone else's system, because it separates the three failures that make a cleaned file look right and total wrong.

The test file was a 60-row vendor expense export: four date formats, three ways of writing currency, 17 vendor spellings that collapse to 11 real vendors, four duplicate invoices and two impossible dates. An AI asked to "clean this and total it" produced a single number, 21,748,113.34, that mixed won into dollars. The real dollar total is 51,464.59. That is an overstatement of 21,696,648.75, and no error message appeared anywhere.

Two more failures came out of the same file. Comparing whole rows as text removed none of the four duplicate invoices, because a trailing space makes two identical rows different strings; keying on the invoice number removed all four and cut USD by 238.75 and KRW by 2,480,000. And a single date-format guess parsed 19 of 60 rows, while DATEVALUE() on a Korean-locale sheet read only 25 of 58 dates even after the values had been separated properly.

The method that survived is unglamorous: put currency in its own column before converting anything, feed the parser an explicit list of formats rather than one guess, de-duplicate on the invoice number, and keep a checksum block on the same sheet so the next person can see whether the totals moved. The guide with every step, the screenshots of the failures and the prompt I used is How to Clean Messy Excel and CSV Data With AI Without Breaking Your Totals.

Run it yourself: the 60-row test CSV is public and needs no sign-in — download the test file (3.4 KB, Google Drive). The vendors, invoices and amounts are invented. The answer key is in the section below.

Which guides cover documents and pasted AI output?

The paste guide answers a smaller question than it looks: when you copy a table out of an AI chat, which of the four available paste paths gives you a real table in Word or Google Docs, and which one gives you fifty pipe characters.

One answer, copied once, pasted four ways on November 15, 2026. The chat app's Copy button followed by Ctrl+V in Google Docs produced a 9×4 table with all 36 cells filled and three real headings. The same clipboard pasted as plain text produced 50 pipe characters, 2 asterisks, 8 backticks and no table at all — 34 unformatted paragraphs. Turning on Markdown in Docs did not rescue that path. Converting the markdown to a .docx with Pandoc 3.9 gave heading levels 1, 4 and 1 and a 36-cell table, but zero cell borders and no cell alignment, because the converter's default table style defines one bottom rule and nothing else.

That last detail is the reason the guide exists. "It pasted as a table" and "it will print as a table" are different results, and the second one is what your reader sees. The full comparison, including the reference-document fix that carries a table style into every future conversion, is How to Paste AI Markdown Tables Into Word and Google Docs Without Broken Borders.

Where does the room stand on PDFs?

There is no PDF guide here yet, and that is the honest gap in this hub. Two PDF jobs I wanted to write up failed my own test, so they stayed unpublished rather than turning into advice I could not stand behind.

The first was turning a styled PDF into an editable slide deck. The reproduction never settled whether the objects survive the conversion or arrive as one flat picture, and a guide that cannot say which would send readers into a wasted afternoon. The second was pulling a scanned table into a spreadsheet: I ran the same 25 cells through two extraction paths and got 25 of 25 on both, which compares the paths but says nothing about whether the numbers are right — the comparison was too clean to be evidence.

A third candidate, checking whether the page numbers and citations in a long PDF actually line up, never got far enough to measure. So the PDF section of this room is a promise with a date on it, not a result. When one of those tests finally produces a number that holds, that guide will be listed here first, and the label Workflow Hub will carry it the same way it carries these nine.

What does exist for PDF-adjacent work today is the paste guide: an AI answer that quotes a PDF, pasted into a document, runs straight into the same table-and-heading problem described above.

Which guides cover schedules, briefs and what you feed the AI?

Two guides deal with the information going in and out of your week, and both found the same shape of error: a plausible answer built on a default nobody chose.

New York to Seoul Meeting Times After DST: 2026 is the scheduling one. Seoul sits 14 hours ahead of New York after the November 1, 2026 clock change, and 13 hours ahead in the October rows of the same table. That single hour splits a recurring invitation into two different meetings: keep 9:00 AM in New York and the Seoul occurrence moves from 10:00 PM to 11:00 PM; keep 10:00 PM in Seoul and the New York occurrence moves from 9:00 AM to 8:00 AM. The guide's point is that you have to choose which city's local time stays fixed, and say so to the organizer, before the series drifts.

TrendRadar AI: Start With a Smaller News Brief is the reading one. TrendRadar is an open-source news and trend monitor that combines trending lists with RSS feeds you pick, and the temptation is to switch on everything at once. In a test on nine invented English headlines, a plain keyword rule for "AI" also matched rail and training, because plain keywords match parts of words. Whole-word rules plus one exclusion fixed the sample. The guide also separates three things people assume arrive together: collection, AI analysis and delivery each need their own setup, and a report saved locally is not a message on your phone.

Which guides cover checking what the AI tells you?

These two are about the moment after the answer arrives, when it reads well and you still do not know whether to believe it.

REA for AI Agents: What It Can Prove About an App uses a tiny JavaScript example to show what an inspection tool can and cannot settle. The tool found the file connection and reported no parsing failures, and left one return value unresolved. All three results were useful, and only one of them was an answer. The habit the guide teaches is to ask for a feature and a question before choosing a tool, because "explain this whole app" produces a report nobody can check.

AI Design Slop: Fix a Generic Page With a Checkable Brief does the same for a generated web page. Two versions of one fictional meeting-room page carry identical room facts, so the only variables left are structure, wording and the primary action. The brief that made the review possible is a single sentence — this page helps this person choose this thing using this information — and the checks that follow are concrete: text contrast, whether the primary button says what happens next, and how the page behaves at phone width.

Which guides cover the habits around the work?

Three guides are about the person rather than the file, and all three came out of a problem I had rather than a tool I wanted to try.

How to Learn With AI: A 6-Step Method From a Real Session came from notes I had saved and could not explain. The method replaces "explain this topic" with six fixed steps per idea — one-line definition, a corrected guess, an analogy from your own job, three core ideas, an application, and one scenario question you answer in your own words. I worked through six project-management concepts in one evening, and every one of my six answers had a gap. The gaps were the useful part.

How to Gamify Boring Tasks: A 15-Minute Mission Template takes a mechanism from the running game Zombies, Run! and turns it into six lines on a card: purpose, finish line, first move, a 15-minute limit, evidence, next move. The guide is careful about the evidence behind it — a 2021 qualitative study of 30 users in 13 countries explains why the game holds people's attention, and no study shows the same trick makes office work faster.

How I Make Claude Code and Codex Share One Worklog is the one that keeps two assistants from overwriting each other's context. Both vendors document the same starting point: each session begins with a fresh window, so nothing the other tool did is visible unless it was written down where both can read it. The fix is a log tree by year, month and day with one file per record: read the last three days before a medium task, never edit a record, and when the log and the files disagree, the files win.

How do the nine guides compare at a glance?

Read down the left column until you recognise the job in front of you, then take the middle column.

The nine guides, by the job they settle (2026-11-15)

What is in front of you	Guide to start with	What it settles
A spreadsheet exported by someone else's system	How to Clean Messy Excel and CSV Data With AI…	Which currency, date-format and duplicate rows move the total, and the checksum that proves the cleaned copy
An AI answer that has to become a Word or Docs file	How to Paste AI Markdown Tables Into Word and Google Docs…	Which of four paste paths keeps headings and all table cells, and what the converter does to borders
A recurring call between two time zones	New York to Seoul Meeting Times After DST: 2026	Whether the New York or the Seoul hour stays fixed, and both local times either side of the clock change
Six industry sites you check every morning	TrendRadar AI: Start With a Smaller News Brief	How plain keyword rules over- and under-match, and which parts of the setup are separate jobs
A generated page that looks finished but does not work	AI Design Slop: Fix a Generic Page With a Checkable Brief	The one-sentence brief and the contrast, control and phone-width checks that make a redesign reviewable
A feature inside an app you did not write	REA for AI Agents: What It Can Prove About an App	What an inspection can point to, and which values stay unknown until the app is running
A task you have postponed for a week	How to Gamify Boring Tasks: A 15-Minute Mission Template	The six parts of a mission card and why the 15-minute boundary does the work
Tutorials you read but cannot explain back	How to Learn With AI: A 6-Step Method From a Real Session	The six-step loop and the closing question that finds what you missed
Two AI assistants editing one folder	How I Make Claude Code and Codex Share One Worklog	The log layout, the read-before and write-after rule, and why the files outrank the log

Which reading path fits today?

Four paths cover almost every reason someone lands on this hub; the rest of the room is best read when the problem shows up.

The file is a spreadsheet and the deadline is this afternoon. Start with the cleanup guide, download the test CSV below, and run your own tool on it before you run it on the real export. You are looking for whether the tool keeps currencies apart and whether it can tell you what it removed.

The deliverable is a document somebody will print. Read the paste guide first and check the borders, not just the cell count. A table that arrives intact and prints without grid lines is the exact failure the guide was written about.

The AI is doing the work and you need to check the result. The app inspection guide and the design review both train the same reflex: ask what evidence would settle this, then look for that evidence rather than a confident paragraph.

The recurring cost is your own attention. The news brief guide reduces what you read, the worklog guide stops you repeating what you already did, and the mission card gets the postponed item moving for 15 minutes.

How do I test the spreadsheet method on the same file?

Download the file, ask your AI to clean it, and compare its numbers against the answer key below — that is the whole test, and it takes about ten minutes.

The file has five columns (Date, Vendor, Invoice, Category, Amount) and 60 data rows. It carries four date formats, amounts written with $, with ₩, with KRW and with no symbol at all, four duplicated invoices and two impossible dates. Every vendor and invoice number is invented for the test, so nothing in it belongs to a real company.

The steps: open it in Excel or Google Sheets, let your AI clean and total it without further instruction, then check four things — how many rows survived, whether the dollar and won totals are reported separately, how many duplicates it removed, and whether the checksum block matches the key. If it reports one combined number, you have already reproduced the failure.

Download: vendor_expenses_60rows_test.csv (3.4 KB, public Google Drive link, no sign-in). The same file and every step are in the cleanup guide.

Answer key for the 60-row test file (measured 2026-11-15)

Data rows in the file — 60
Rows with a usable date and amount — 58 (36 USD, 18 KRW)
Unique rows after de-duplicating on invoice number — 54
Duplicate invoices — 4 — INV-1003, INV-1008, INV-1021, INV-1034
Correct USD total — 51,464.59
Correct KRW total — 19,128,000
Money removed by de-duplication — USD 238.75 and KRW 2,480,000
What a currency-blind cleanup reports — 21,748,113.34 (overstated by 21,696,648.75)

What does Charrette AI Room publish next?

The room grows one guide at a time, and a new one joins the section above that matches its job rather than sitting at the top of a feed.

Next in line are the PDF tests that have not yet produced a number worth publishing, then a scanned-document comparison that measures whether an extraction is right rather than merely consistent. A guide lands here only when it can name the file, the date and the check at the end; if a test refuses to settle, the note stays in the drawer and the section stays short.

Every published guide as of November 15, 2026:

How to Paste AI Markdown Tables Into Word and Google Docs Without Broken Borders
How to Clean Messy Excel and CSV Data With AI Without Breaking Your Totals
How to Learn With AI: A 6-Step Method From a Real Session
New York to Seoul Meeting Times After DST: 2026
TrendRadar AI: Start With a Smaller News Brief
AI Design Slop: Fix a Generic Page With a Checkable Brief
REA for AI Agents: What It Can Prove About an App
How to Gamify Boring Tasks: A 15-Minute Mission Template
How I Make Claude Code and Codex Share One Worklog

Written and edited by JH-A. The tests described here were run between October 22 and November 15, 2026. Questions about a test, a file or a correction are welcome on the contact page; how this site handles sources and mistakes is on the editorial policy page.

Frequently Asked Questions

Do these workflows need a paid AI subscription?

Most of them do not. The spreadsheet cleanup guide is built on formulas you run in the sheet, and the paste guide is about the clipboard in Word and Google Docs plus the free Pandoc converter. One guide, the six-step learning method, comes from a session on a paid Claude Max plan, and it says so where the method depends on it.

Can I download the file used in the spreadsheet guide?

Yes. The 60-row CSV is public on Google Drive with no sign-in needed: 3.4 KB, invented vendors and invoices, and the answer key values are printed both in the guide and in the table above. Compare your own tool's cleanup against them before you trust it with a real export.

Are these guides reviews that rank AI tools?

No. I do not run model-versus-model benchmarks, so nothing here says which AI is best. Each guide takes one file, one job and one date, measures what happened, and states what it did not measure. Where a result came from a constructed example, the guide says so — the design review and the app-inspection walkthrough are both constructed.

How often is this hub updated?

Every time a guide is published. This page was assembled on November 15, 2026 and lists nine guides; a new one joins the section that matches its job, and the PDF gap described above is the first section that will change.

How to Paste AI Markdown Tables Into Word and Google Docs Without Broken Borders

JH-A — Sun, 15 Nov 2026 10:43:50 +0000

Copying a table out of an AI chat gives you two different things at once, and which one your paste uses decides whether you get a real table or fifty pipe characters.

In a test run on 2026-11-15, the chat app's Copy button followed by a normal paste in Google Docs produced a table object with 9 rows, 4 columns and 36 filled cells, and left 0 markdown characters behind. The same clipboard pasted as plain text produced no table at all and 50 pipes. On the Word side, Pandoc 3.9 produced real Heading 1, 2 and 3 styles and a 36-cell table in 1.59 seconds, but the file it wrote carries no cell grid lines and dropped every column alignment marker.

One Copy click puts two versions on the clipboard

The chat app renders the markdown you see, and its Copy button puts both the rendered version (as HTML) and the source version (as text) on the clipboard. The application you paste into then picks one.

Clipboard format	What it holds	What a paste does with it
HTML	Rendered tags: heading elements, a table with rows and cells, bold and code spans	Google Docs built 1 table, 3 heading paragraphs and 7 bullet paragraphs; 0 markdown characters arrived
Plain text	The markdown source: hash marks, pipes, asterisks, backticks, one separator row of dashes	Google Docs inserted 34 plain paragraphs; 50 pipes, 2 asterisks, 8 backticks and 3 hash-marked lines arrived as characters

The plain-text version is not a degraded copy of the answer. It is the answer exactly as the model wrote it, which is what you want when the next stop is another tool, and the wrong version when the next stop is a person reading a document.

The four paste paths, measured

Path	What arrived	What did not
Copy button, then Ctrl+V in Google Docs	3 heading paragraphs (2 at level 2, 1 at level 3), one 9×4 table with 36 of 36 cells filled, 7 bullet paragraphs at 2 levels, bold and monospace text	The code block: its fence and language label arrived as a plain "Python" line, and the code as monospace paragraphs
Copy button, then Ctrl+Shift+V in Google Docs	All 34 paragraphs of text, every word intact, 1,566 characters	0 headings, 0 tables, 0 bullet objects, 0 bold runs; 50 pipes and 2 asterisks stayed in the text
Google Docs "Paste from Markdown"	Not measured: the command opened a dialog asking to install a Chrome extension	Alignment, footnotes and code blocks on that path
Pandoc GFM to DOCX	Heading 1 ×1, Heading 2 ×4, Heading 3 ×1 as real styles, 9×4 table with 36 of 36 cells, 7 bold runs, 1 hyperlink, 1 footnote, 12 list paragraphs at 2 levels	Cell grid lines (the table style defines one rule under the header row), 0 aligned cells

The rich paste is the only path that gave a real table with no cleanup. Pandoc is the only path that produced Word heading styles, a working hyperlink and a real footnote.

Turning on Markdown in Google Docs did not fix a plain paste

Google documents two different things and they are easy to conflate: typing markdown as you write, and converting markdown on paste. I tested the second with the first switched on.

Tools → Preferences → Enable Markdown, confirmed on.
Ctrl+Shift+V with the same markdown clipboard: 34 plain paragraphs again, 50 pipes, 2 asterisks, 8 backticks. Identical to the run with Markdown off.
"Paste from Markdown" under the Edit menu and the right-click menu both opened a dialog titled "Turn on copy, cut and paste", asking for a Chrome extension.

So that result says more about this browser than about Google Docs. What a measured number does support is narrower: on this setup the setting alone did not change what a text-only paste produced.

The one-line Pandoc command

pandoc -f gfm -t docx -o report.docx report.md

On a 536-word markdown report with a 4-column table, a footnote and a link:

1 H1, 4 H2, 1 H3 in the source → Heading 1 ×1, Heading 2 ×4, Heading 3 ×1 as Word paragraph styles an outline can read
4-column table, header row plus 8 data rows → 9 rows × 4 columns, 36 cells, 36 filled, header row marked to repeat across page breaks
Alignment markers :--- and ---: → 0 cells carry an alignment property
1 link → 1 hyperlink relationship, not a pasted URL
1 footnote → 1 footnote in a real footnote part

The alignment loss is not the markdown's fault. Converted to HTML first, all 36 cells carried their alignment as an inline style, 18 left and 18 right; the property disappeared in the DOCX step. The border result is the same kind of fact: nothing in the file draws cell grid lines. Both are one minute of work in Word (select the table, apply All Borders), and if you convert the same documents often, set the look once in a reference document and pass it with --reference-doc.

Which path to use

If you need	Use	Why, from the numbers
A table that looks right in Google Docs now	Copy button, then Ctrl+V	36 of 36 cells and 3 heading styles arrived in one paste, with 0 markdown characters left
The AI's exact markdown, for another tool	Copy button, then Ctrl+Shift+V, or Copy as markdown	The source arrived unchanged: 1,566 characters including the 50 pipes a parser needs
A Word file with outline levels, a link and a footnote	`pandoc -f gfm -t docx`	Heading 1–3 as real styles, 1 hyperlink, 1 footnote, 36 of 36 cells
A Word file you will not touch afterwards	Pandoc, plus a table style applied once	Borders and alignment are the two things the default output lacks

The pattern is worth memorising: rich paste for people, plain paste for machines, a converter when the destination is a file.

The full guide, with the test conditions, the sample sizes and the document-API counts behind every number above, is on my blog: Read the full guide and test benchmarks on Charrette AI Room.

Originally published at https://c-roomm.blogspot.com/2026/11/paste-ai-markdown-tables-word-google-docs.html