The 3 gaps
Three clocks on every item. Type the four times and the sheet fills the hours.
- 01MAKE = drafted minus asked. The part AI speeds up.
- 02WAIT = checked minus drafted. Where the jam usually is.
- 03SHIP = sent minus checked. Formatting, uploading, chasing.
If WAIT beats MAKE, AI isn't your problem.
How to log
Columns: Item, Asked, Drafted, Checked, Sent, Sent back (Y/N), Why sent back. The sheet then fills MAKE hrs, WAIT hrs, SHIP hrs, Total hrs and Biggest gap. Formulas run for 100 rows.
Times: type them like 24/11/2026 09:10 so the sheet can do the maths. MAKE, WAIT and SHIP are in hours.
- 01Pick ONE job AI touches (10+ times a fortnight, AI drafts it, a person checks it, someone outside waits on it).
- 02One row per item. Type the real date and time at each step as it happens. Do not back-fill from memory.
- 03Asked = when the request came in. Drafted = first draft exists. Checked = a person finished checking it. Sent = it left the building.
- 04Sent back? = Y if the check sent it back for fixes. Write why in a few words.
- 05After two weeks, look at the Read-out tab, or paste the Log into Claude or ChatGPT with readout-prompt.txt.
- 06Change ONE thing at the hot spot (the fix menu). Log the next fortnight. Did the hot gap drop?
The read-out prompt
Paste your log into Claude or ChatGPT with this prompt. It gives you the numbers and patterns, not advice. If a row is missing a time, the prompt tells it to list that row and leave it out of the maths.
Here is a 2-week log for one job. Columns: item, asked, drafted, checked, sent, sent back (Y/N + why). For every item, work out in hours: MAKE = drafted - asked WAIT = checked - drafted SHIP = sent - checked Then tell me: 1. The median of each gap 2. Which gap is biggest 3. The 3 items that waited longest and what they share (day of week, type, who checked it, time of day) 4. How many were sent back, grouped by reason Numbers and patterns only. No advice yet. If a row is missing a time, list it and leave it out of the maths. Show me a small table first, then the four answers. [paste the log here]
Fix menu
Change one thing. Log the next fortnight. Check if the hot gap dropped.
WAIT is biggest (drafted to checked)
The work is done and sitting there. This is the jam the Harvard paper found.
- 01Book two 20-minute check slots a day in the calendar. Same times every day. "When I get to it" is the jam.
- 02One named checker per job. Shared ownership means nobody checks.
- 03Cap the pile: if more than 5 drafts are waiting, stop making new ones until the pile is under 5.
- 04Check in batches of the same type. Five proposals in a row is faster than five different jobs.
Lots sent back (high sent-back rate)
The check is doing the thinking that should happen before the job.
- 01Write 5 pass/fail lines before the job starts (use check-template.md).
- 02Give them to the AI with the brief and tell it to check its own draft against each line, with evidence, before it hands over.
- 03Keep the check in its own file the AI can read but doesn't edit. (Same idea as the Claude Code team's verification write-up: the work shouldn't be able to change the test that judges it.)
- 04Every time something gets sent back for a new reason, add a line. Delete lines that never fail after a month.
SHIP is biggest (checked to sent)
The work is approved and still not out the door.
- 01Template the handoff: same format, same place, same send step every time.
- 02Write down the send steps once. If it's more than 5 steps, that's your next automation.
- 03Send at the moment of approval, not "later today".
MAKE is biggest (asked to drafted)
Good news. That's the gap AI actually shrinks.
- 01Move the first draft to AI if it isn't already.
- 02Save the brief that works as a template so every request starts the same way.
- 03Then scan the next job.
Don't
Don't fix a jam with more drafts. More output makes the pile bigger.
Don't report "we made 40 this month". Report how many finished and how long they waited.
Check template
Copy, fill, and keep it in its own file. The AI reads it. It doesn't edit it. A draft passes only if every line is a clear yes.
# CHECK: [job name] Owner: [one person] Last changed: [date] Rule: this file changes only when a draft fails for a new reason. The AI reads it. It doesn't edit it. A draft passes only if every line is a clear yes. 1. [ ] Facts: every number, name and date matches [source of truth, e.g. the CRM, the price list]. 2. [ ] Ask: it answers the exact request in the brief (quote the request, quote the answer). 3. [ ] Voice: it sounds like us. No [banned words / phrases]. 4. [ ] Format: [length, layout, file type, where it goes]. 5. [ ] Next step: it ends with one clear next step for the reader. ## Prompt to use with it Before you hand this draft to me, check it against every line in CHECK.md. For each line, answer PASS or FAIL and quote the part of the draft that proves it. If any line fails, fix the draft and check again. Only hand it over when all five pass. Don't change CHECK.md. ## Sent-back log (add a line when a new reason shows up) | Date | Item | Why it was sent back | Line added? | |---|---|---|---|
Prompt to use with it:
Before you hand this draft to me, check it against every line in CHECK.md. For each line, answer PASS or FAIL and quote the part of the draft that proves it. If any line fails, fix the draft and check again. Only hand it over when all five pass. Don't change CHECK.md.
The 2-week schedule
Day 0 (today, 15 min)
- 01Pick one job. Write it at the top of the sheet.
- 02Tell whoever does the job: "For two weeks, type the time at each step. That's it."
Days 1 to 10 (every workday, about 2 min)
- 01One row per item. Type Asked, Drafted, Checked, Sent as they happen.
- 02Sent back? Y or N, plus a few words on why.
Day 14, second Friday (15 min)
- 01Open the Read-out tab, or paste the Log into Claude or ChatGPT with readout-prompt.txt.
- 02Find the hot spot (biggest median gap).
- 03Pick ONE fix from fix-menu.md. Put it in the calendar.
Days 15 to 28
- 01Keep logging. Same job, same columns.
Day 28, fourth Friday (15 min)
- 01Compare: did the hot gap drop? Did the sent-back rate drop?
- 02Yes: keep the fix and scan the next job. No: try the next fix on the menu.
Want the wait taken off the team, not just measured? Innoval OS builds AI workflows for service businesses.
Talk to JoelSource
Not peer reviewed yet. Software teams only. The 14% is only borderline significant. Innoval OS, Brisbane, Australia-wide.