Imashime

npx add-mcp 'https://imashime.waitlab.workers.dev/mcp'

Imashime

Notes on mistakes AI assistants made in everyday work, and how to ask next time

You asked an AI to do something, and it did something else. Imashime collects those moments: 54 short notes from writing, spreadsheets, research, office work and computer maintenance. Each note gives the situation, what the AI did, an example of how to ask next time, one question to ask, and a way to check for the mistake yourself. Nine of the checks use a terminal command.

Imashime (ee-mah-shee-meh) is Japanese for a warning you keep in mind so you don't repeat a mistake.

日本語

How to use it

Before you hand a task to an AI, find its category below and skim the titles. If one fits, copy the "How to ask next time" example into your request and change it to match your task. The "Check it yourself" line on each note shows one way to catch the mistake; some checks belong before the AI starts, others after it says it's done.

You can also let your AI look here before it starts (see "Connect your AI"). It reads the notes as background, not as commands to obey.

Where the notes come from

The site is run by one person in Japan, under the name WaitLab. Most notes come from that person's own work with Claude Code between May and October 2026, with the original task replaced by a different one so that no client or private details show. Eight notes draw on videos by Yuru Computer Science Radio, a Japanese YouTube channel, and one of them comes from a video alone. Each note names its source.

Each note was first written in Japanese. The English version keeps the same situation, mistake and question, and adds the "Check it yourself" line, which the Japanese note does not have. Where a scene depends on life in Japan, the note says so. Every note also has a Japanese page; the link is at the bottom of each note.

Categories

General (any task) (7)

• Checked only the part it added, then reported "done"
• Treated "not in the places it searched" as "doesn't exist"
• Tested in a shrunk desktop window, not a phone, and nearly called it a bug
• Opened the file in another app and said "all OK" without comparing the results
• Copied the number of items from a reference, not the reason for it
• Fixed each reported problem one at a time and said "fixed" every time
• Didn't account for the cleanup, or the time the person couldn't use the computer

Writing (12)

• Brought in another document's rules and gave advice it had no standing to give
• Rewrote a whole long document and duplicated part of it
• Wrote an example that fits only some of the readers
• Told readers something was ready when that was their call
• Used phrases that make sense but nobody actually says
• Made the part the AI found interesting the focus for readers
• Wrote "it couldn't go further," but another way worked
• The title was stronger than the source's conclusion
• A quote was turned into a flat statement
• Swapped in new sentences that no longer fit the ones around them
• Cut words to make it shorter until the point no longer came across
• Presented an idea common abroad as if it were common knowledge in Japan

Computer maintenance (7)

• Guessed why a scheduled job didn't run, without reading the logs
• Concluded "there's nothing left to delete" without looking
• Called an estimate "measured," and it was far off
• Deleted the original file before making a backup
• Fixed one cause and called it done without measuring
• Kept asking a person to run a check the AI could have run itself
• Picked the app to quit by part of its name and quit unrelated apps too

Office work (6)

• Nearly scaled up a single timed run into an estimate
• Accepted "a person enters this by hand" in a procedure without asking why
• Wrote a rule without saying where it applies
• Mixed a separate problem it found into the requested work
• Filled in a cut-off instruction by guessing, and kept going
• Handed the work to another AI and reported "done"

Data cleanup (5)

• Split one client into two companies but left the contacts as they were
• Dropped rows vanished from the denominator and the result became "100%"
• Read "zero problems found" as "no problems"
• Matched on the wrong key and nearly reported a bug that wasn't there
• Treated "couldn't measure" the same as "failed"

Making documents (5)

• Made a value changeable in one place, but missed spots showing the old one
• Asked a person to choose between options built on a broken assumption
• Wrote only the on-screen steps and missed how the person actually works
• Called it a bug based on a mental picture of "how it's normally done"
• Checked only the first page of a multi-page document

Research (4)

• Wrote a service's limits from memory
• Tried one method, then decided "it can't be done"
• Built a plan on an assumption marked "unconfirmed"
• Pulled out only the requested topic from long material and threw away the rest

Spreadsheets (4)

• Nearly brushed it off with "it was already wrong before"
• Put a note where readers would take it for a note on a different number
• Read the last number in a table, used the next one, and created a duplicate
• The breakdown didn't add up to the total

Scheduling and messages (3)

• Showed one venue name the same way in two schedules for different readers
• Moved on without noticing the reply had already come
• Picked "next Monday," and it was a holiday

Video editing (1)

• Cut laughter as if it were "time when nobody is talking"

Connect your AI

No login. In Claude Code, run:

claude mcp add --transport http imashime \
  https://imashime.waitlab.workers.dev/mcp

In claude.ai or Claude Desktop, add it from the connectors directory.

Two tools come with it.

search_lessons (read-only) takes keywords and an optional category and returns up to three notes. English keywords return the English text.

submit_lesson sends a story of an AI mistake to the site's author. Your AI is told to use it only when you ask, and only with text you have approved. Before it is stored, fixed patterns remove URLs, code, key-like strings and similar parts, and mask personal details such as email addresses and phone numbers (not names). The patterns are imperfect, so they may remove too much or miss something; your AI gets the stored text back so you can check it. The text is never published, is read by the operator with Claude (Anthropic), may be used to train Anthropic's models, and is deleted from the site's storage after 90 days.

All 54 notes are also in lessons.json.

Help

Questions, corrections and removal requests go through the contact form. The form is in Japanese, but the top of the page explains it in English, and you can write in English.

Privacy policy and Terms: an English summary is at the top of each; the Japanese text below it is the governing version.