The three findings that decided the design

1. Your instinct is right, and it is a real improvement on its source

The question echoes Jake Knapp's "Highlight" from Make Time, but his version is prospective and daily: "what do I want the highlight of my day to be?" Yours is retrospective and weekly. That mutation does a second job his does not: it selects one specific moment, which is exactly what every deeper reflection model needs and can never get from "describe your week."

Putting it first is also what people who run reflection professionally do. The VA's adaptation of the Army's After Action Review takes the Army's single evaluation question and deliberately splits it in two so that "what went well, and why?" comes before "what can be improved?" That change was made by clinicians running these sessions for real. Tiago Forte opens his own reviews with gratitude and names the mechanism better than anyone:

"It puts me in an expansive, generative, abundance-oriented state of mind that makes the following questions more like gentle inquiries, rather than a harsh interrogation."

Honest caveat: no experiment has actually tested whether asking about wins first improves the answers that follow. What exists is a convergent case: facilitators do it deliberately, positive affect measurably broadens thinking and improves creative problem solving in the lab (Fredrickson 2005; Isen, Daubman and Nowicki 1987), and naming small wins is itself the strongest predictor of a good working week (Amabile's 12,000-entry diary study). Plausible and evidenced at each link, untested as a chain. It costs nothing to do, so we do it.

2. One question at a time genuinely beats a form, and this is measured

Xiao et al. (ACM TOCHI 2020) ran roughly 600 participants and 5,200 free-text responses. Half filled in a standard survey form; half talked to a bot that asked open questions one at a time and probed the answers. The conversational condition produced significantly better responses on informativeness, relevance, specificity and clarity, plus significantly more disclosure.

Kim et al. (CHI 2019) found something sharper still. A casual, warm tone improved answer quality only in the conversational format. In the form condition, tone made no measurable difference at all. So talking to you like a friend who has been paying attention is not decoration on this skill, it is the active ingredient.

3. Asking is not enough. Asking informed is what works

Park et al. ("Thinking Assistants") tested four conditions: a blank box, questions only, advice only, and informed inquiry, meaning questions grounded in real knowledge of the person's actual situation. Informed inquiry beat all three. Users disclosed significantly more than in the advice condition, and complained about length far less.

Which answers the question I asked you earlier. A skill that asks "what was the highlight of your week?" cold is the weaker questions-only condition. A skill that reads the git logs, the session summaries and last week's check-in first, and then asks, is doing the thing that actually works. So yes, Claude reads your week before asking. But it never makes you recite it.

---

The problem nobody else has solved: 17 projects

Reviewing a project portfolio is where every weekly review either does real work or collapses. As a checklist it is grinding tedium. As a conversation it is seventeen questions nobody sits through. Either way it reliably triggers what practitioners call the guilt spiral:

"You open your project list and see three projects with no progress in a month. The guilt spiral begins: I'm terrible at this. I'll never get caught up. Why bother?"

The resolution is the one thing an AI can do that paper cannot. Claude reads the state of all of them, identifies the two or three that moved and the one or two you named last week and didn't touch, and brings only those. You never enumerate anything. Most weeks most of seventeen projects moved zero, and that is normal, not a failing.

The same logic retires the biggest chunk of GTD. Of its eleven steps, only one is a reflection question; the rest is inventory maintenance an assistant can do silently. Which is worth saying plainly: the thing everyone cites as "the weekly review" is almost entirely filing, and it is the most-abandoned part of the system because it runs one to three hours.

---

Why fifteen minutes, and not thirty

The Wipro experiment is the strongest single piece of evidence in the whole set. Workers given 15 minutes of structured end-of-day reflection scored 22.8% higher on their final assessment than controls, despite the controls having worked 15 minutes longer each day. The mechanism ran through increased self-efficacy. Adding a sharing component produced no significant extra benefit, which means reflecting alone costs you nothing.

One uncomfortable finding worth flagging rather than burying: weekly is a low dose for reflection interventions. A systematic review found 90% of studies with daily or near-daily participation showed benefit, against only 25% of once-weekly studies. The honest reading is that weekly is the right container for portfolio and pattern work, but the wins-and-gratitude part probably wants a lighter, more frequent touch. Worth considering a thirty-second daily version later.

---

The best 22 questions found

The design uses about six of these per session, drilled twice each, rather than all 22 asked once. Depth comes from repeating one question, not from breadth. The Coaching Habit's four-minute drill is the model: ask the challenge question, then "and what else?", then "and what else?", then ask the challenge question again.

---

What killed the other implementations

Worth knowing, because these are the specific ways this could fail:

• Prompt repetition. The Reflectly failure, and the one most likely to bite something you run 52 times a year. "What made you smile today?" is beautiful in week one and dead by week six, because you know your answer before you open it. The fix: the six beats stay fixed, but the wording comes from what actually happened this week, never from a static list.
• Length. Every framework that survived is shorter than GTD; every one that died got longer. One skill author put it well: a 90-minute review gets skipped by week three.
• Streaks. They run on loss aversion, so the mechanism that makes them satisfying is the same one that makes breaking one feel like total failure. Roughly half of habit-app users are gone by day 60. Habit research (Lally et al.) found missing a single occurrence does not measurably damage habit formation, so the streak is lying to you. This skill has no counter. It has a "never miss twice" rule instead.
• Forced positivity. Appreciative Inquiry gets criticized for suppressing real problems, and AI journaling apps get accused of training people to "rehearse emotional platitudes." A check-in that cannot hear "this week was bad" gets avoided in exactly the weeks it matters. The fix is the coping questions from solution-focused therapy, which stay strengths-based without requiring the week to have been good.
• Creepiness. About 21% of people in one study found AI probing on personal ground uncomfortable. Declining a probe has to be frictionless.
• Advice dumping. The opposite failure. In the Thinking Assistants study, 8 of 20 people in the advice-heavy condition complained the responses were long and overloaded and gave no actual guidance on how to reflect.

---

The design, in one screen

Six beats, 15 minutes, hard ceiling.

1. The highlight (3 min). The opener, always. Then "and what else?" twice.
2. What moved (3 min). Specific, because Claude read the actual evidence. Wins first, with why.
3. What didn't (3 min). One or two exceptions only. Plan problem or execution problem? And explicit permission to kill things.
4. Patterns (2 min). The thing paper cannot do. One bad week is noise, three is a signal.
5. Next week (3 min). Three priorities, not five. Each with a "done means what exactly" and a Monday first action. Then: what are you saying no to?
6. Close (1 min). Energy 0 to 10 as data, not therapy. Then "what was most useful about this?", every single time, because that is how the skill gets tuned to you.

Plus a ten-minute crisis version (highlight, the one thing that didn't move, next week's three) because a short version you actually run beats a full version you skip.

The tone contract is two borrowed lines: the Army AAR's "there will be no searches for the guilty" and Ryder Carroll's "transform any guilt into curiosity."

---

Sources

Frameworks: GTD official weekly review checklist · Make Time, Choose a Highlight · Guide to the After Action Review (VA) · The US Army's After Action Reviews · Bullet Journal, Reflection · Forte, The Design of a Weekly Review · The Coaching Habit, seven questions · Solution-Focused Therapy Institute · FSG Guide to Appreciative Inquiry · Cambridge Reflective Practice Toolkit (Gibbs)

Research: Xiao et al., chatbot versus survey response quality · Kim et al., conversational style and data quality · Park et al., Thinking Assistants · Di Stefano et al., Learning by Thinking (Wipro) · Lally et al., habit formation · Amabile and Kramer, the progress principle · Fredrickson and Branigan, broaden and build

Implementations studied: TheCraigHewitt/skills, ceo/weekly-review · diary-planner-plugin (the Goals Graveyard) · Sunsama Weekly Review · Stoic's on-device personalized prompts

Two widely circulated GTD statistics were found to be unsourced or likely fabricated and were deliberately excluded. Nothing on this page rests on them.