In a randomised trial, developers using AI felt 20% faster and were measurably 19% slower. Twenty-three prompts biased toward things that leave a trace rather than a feeling.
You cannot feel productivity
In July 2025, METR ran a randomised controlled trial with 16 experienced open-source developers across 246 tasks in repositories they knew well. Before starting, they forecast that AI tooling would make them 24% faster. Afterwards, having done the work, they estimated they had been 20% faster.
Measured against the clock, they were 19% slower.
The study is narrow and should be described as such — experienced developers, mature codebases, a particular set of tools at a particular moment. It does not generalise to all work. But the perception gap is the finding, and it is the one that matters here: everyone involved sincerely believed they were faster, and the belief was wrong by nearly forty points.
Which is the uncomfortable premise of this entire pack. Feeling productive and being productive are separate variables, and the feeling is generated by effort, momentum and novelty rather than by output. Any productivity system built on how a day felt is built on a signal with no demonstrated relationship to results.
So the prompts here are biased toward things that leave a trace: written plans, explicit priorities, recorded decisions, audited time. Not because artifacts are virtuous, but because they are checkable and a feeling is not.
The other half of the problem is structural
In No Task Left Behind?, Gloria Mark, Victor González and Justin Harris observed 24 information workers in detail and found work to be, in their words, highly fragmented
: people average little time in working spheres before switching and **57% of their working spheres are interrupted**.
When work is interrupted, it is usually resumed the same day — but **more than two intervening activities occur before it is**.
That last clause is the one worth sitting with. The cost of an interruption is not the interruption. It is the two unrelated things you do before you find your way back.
A note on a number you have almost certainly seen: the famous 23 minutes and 15 seconds
to recover from an interruption does not appear in this paper, though it is cited to it constantly. It comes from Mark's June 2006 interview in the Gallup Business Journal, where she said interrupted work was resumed the same day 81.9% of the time and, on average, in 23 minutes and 15 seconds. Real figure, real researcher, wrong citation — and a fair illustration of why the research pack insists you follow a statistic back to its source before you repeat it.
Between the two studies: your sense of your own throughput is unreliable, and your day is more fragmented than you think. Everything below is a response to one or the other.
Decide what deserves the time
Prioritize Your Task List is the entry point, and its value is forcing the trade-off into the open. A list where everything is important is a list that has not been prioritised; it has been recorded.
Organize a Brain Dump Into an Action Plan is the right first move when the problem is that everything is in your head at once. Getting it out is a prerequisite to judging any of it, and the disorder is usually smaller written down than it felt.
Turn a Vague Goal Into SMART Goals and Break a Big Goal Into an Action Plan handle the gap between an intention and something you could actually start on Monday. Most abandoned goals were never converted into a next action.
Make a Plan to Tackle a Backlog or Overwhelm is for when the list has become the problem. Some of it needs doing, some of it needs deleting, and separating those is most of the relief.
Protect the time you decided to spend
Given 57% interruption rates, the interesting question is not how to work faster but how to get uninterrupted blocks at all.
Plan and Time Block Your Day assigns work to specific hours rather than to a list. The mechanism is that a block has to displace something to be added, so the arithmetic of an over-full day becomes visible in the morning instead of at six in the evening.
Beat Procrastination on a Task starts by diagnosing rather than motivating — procrastination on an ambiguous task, an aversive task, and an overwhelming task are three different problems with three different fixes, and the advice for one is useless for the others.
Audit Where Your Time Goes is the METR finding turned into a practice. It is also the least popular prompt on this page, for the obvious reason: it replaces an impression with a record, and the record is rarely flattering.
Run a Weekly Planning Session and Do a Personal Reflection or Retrospective provide the loop. Without a recurring checkpoint, a system degrades quietly and you notice a quarter later.
Build a Habit That Sticks works on the cue, the friction and the smallest viable version, because the common failure is starting at a size that requires motivation to sustain.
Get the work done
Summarize a Long Document Accurately carries the constraint that makes it usable: it works from the document and does not add. A summary that introduces a claim the source did not make is worse than not reading it, because you now believe something specific and false.
Turn Meeting Notes Into Decisions and Action Items separates what was decided from what was discussed — and marks the things that only sounded decided, which is where most post-meeting confusion originates.
Extract Structured Data From Messy Text and Analyze a CSV or Dataset cover the genuinely large win: converting unstructured material into something you can act on. Check the output — see the analytics pack for why an arithmetically correct number can still be wrong.
Write a Spreadsheet Formula From a Description is a small prompt with an outsized hit rate, because the hard part of a formula is remembering the syntax rather than knowing the logic.
Turn Requirements Into a Structured Spec surfaces the ambiguity in a request before someone builds the wrong thing from it.
Make it repeatable
This is where personal productivity turns into something that outlives your attention.
Create a Checklist or Repeatable Process converts a thing you did well once into a thing that gets done well by default, including by someone else.
Plan a Project From Start to Finish handles work with dependencies, where the failure is discovering in week five that step two needed to happen first.
Map a Process Before You Try to Fix It enforces the ordering that most improvement efforts skip. From operations, along with Choose Software or a Vendor Without a Six-Month Evaluation — because tool evaluation is where a great deal of productive-feeling time disappears.
Work Out Your Real Capacity Before You Promise a Deadline counts the interruptions and overheads that never appear in a plan. Given the fragmentation research, planning against theoretical available hours is planning against a number that does not exist.
Plan a Trip Itinerary is the same planning discipline applied to something enjoyable, and it is here because the pack is not only about work.
Where this stops
None of these prompts do the work, and a well-organised plan is not progress. It is entirely possible to spend a productive-feeling morning restructuring a task list and produce nothing — which is the failure mode this pack is most at risk of enabling, so it is worth naming plainly.
They also work only from what you tell them. They cannot see your calendar, your obligations, or the political reality that makes the obvious plan impossible, and they will not know that a task you described as small is the one that always takes a week.
And they are not a substitute for addressing a cause. Chronic overload is usually a capacity or prioritisation problem at the level of the team, and no personal system fixes an unsustainable workload — see the management pack if you are the person who could change it. Where the difficulty is persistent focus, attention or exhaustion rather than method, that is worth raising with a professional rather than optimising around.
Sources
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025 — 16 experienced developers, 246 tasks in familiar repositories; forecast 24% faster, self-assessed 20% faster, measured 19% slower
- Gloria Mark, Victor M. González and Justin Harris, No Task Left Behind? Examining the Nature of Fragmented Work, CHI 2005, University of California, Irvine — detailed observation of 24 information workers; 57% of working spheres are interrupted, and more than two intervening activities occur before interrupted work is resumed
- Gloria Mark, interviewed by Jennifer Robison, Too Many Interruptions at Work?, Gallup Business Journal, 8 June 2006 — the actual source of the widely misattributed
23 minutes and 15 seconds
figure: interrupted work resumed the same day 81.9% of the time, on average in 23 minutes and 15 seconds