ChatGPT Prompt to Run a Project Retrospective

Trayline delivers 1,400 campus dinners a night across three universities, and its spring dinner pilot in the north residence halls burned 21,000 dollars while on-time completion slid from 94 to 71 percent. Founders Priya Raman and Marco Vidal closed the quarter with three spreadsheets, a Slack channel of driver complaints, and no shared answer to what actually broke. Run from memory, the retro would have settled on “the app felt slow” and adjourned.

The prompt below walks the raw evidence into themes, traces each theme to a root cause, and returns three actions with a single owner and a date. Fill the brackets, paste the whole thing into ChatGPT, and you get a document the team can argue over in forty minutes instead of re-litigating the quarter for three weeks.

How the Dinner Pilot Unraveled

Trayline launched dinner delivery in February to 1,100 opted-in students. Kitchens closed at 8:15pm, couriers were paid per drop, and a single dispatching spreadsheet assigned orders in batches of twelve.

  • Wait times: median delivery hit 63 minutes in March against a 35-minute promise.
  • Refunds: 412 orders refunded, 78 percent of them from the two halls farthest from the main kitchen.
  • Courier churn: 19 of 34 couriers quit or went inactive during the pilot.
  • Sentiment: the shared complaint channel logged 260 messages, most mentioning cold food rather than lateness.

What Trayline lacked was not data. It was a way to turn four conflicting narratives, the routing problem, the kitchen closing time, the pay model, and the cold-food complaints, into a small set of causes that the team could agree to fix. That is the job the prompt does.

The Copy-Paste Retrospective Prompt

Paste the whole block, replace the bracketed fields, and keep the structure intact. The output rules are what stop ChatGPT from producing a vague summary.

You are facilitating a quarterly retrospective for a small startup. Be direct and
evidence-driven. Do not praise the team and do not pad the response.

Context:
- Company: Trayline, a campus meal-delivery startup running dinner across three
  universities.
- Period reviewed: February to April, the north residence hall dinner pilot.
- Goal we set: 35-minute median delivery, completion rate above 90 percent.
- Actual result: 63-minute median delivery, completion 71 percent, 21,000 dollars
  lost, 412 refunds, 19 of 34 couriers churned.

Raw evidence (paste your notes, messages, and metrics here):
[EVIDENCE]

Produce the retrospective in exactly these sections:
1. What we expected vs what happened. Two short paragraphs, numbers only, no adjectives.
2. Themes. Cluster the evidence into at most four themes. For each theme give a
   one-line label, the evidence that supports it, and how many separate incidents point
   to it.
3. Root causes. For each theme, name the underlying cause and separate what we
   controlled from what we did not (weather, campus rules, exam week).
4. What worked. List at most three things worth keeping, with the evidence for each.
5. Actions. Exactly three actions. Each one needs an owner, a due date within 30 days,
   and a metric that proves it worked.
6. Open questions. List what we still cannot answer from this evidence and what data
   would settle it.

Constraints:
- No action may be "communicate better", "improve the process", or "be more careful".
- Every claim must trace to the evidence block or to the metrics above.
- If the evidence contradicts the metrics, say so explicitly.
- Keep the whole document under 700 words.

Why the constraints matter: the three-action cap and the ban on vague verbs are the difference between a retro that ends in a plan and one that ends in a list nobody owns. Trayline’s first draft retro produced nine actions; three survived contact with the calendar.

Fields You Must Replace Before Running It

Swap the bracketed and quoted values for your own run. Everything outside them is structural and can stay.

  • Company and period. Replace “Trayline”, “February to April”, and the hall name with your own team and window.
  • The goal and the actuals. Put your real target and your real result. If the goal moved mid-quarter, say so and give both targets.
  • Metrics. Keep the four or five numbers that the team would argue about. Drop vanity metrics.
  • EVIDENCE. Paste the messy material: complaint messages, refund reasons, courier exit notes, the shared doc. More raw text means better clustering, up to roughly two pages.
  • Action window. Change “within 30 days” if your team ships on a different cadence, but keep the deadline concrete.
  • Word cap. Lower it to 500 for a smaller team; the cap is what forces the model to choose.

What the Output Should Look Like

For Trayline, the returned document named four themes. The one that changed the plan was “distance, not speed, drove cold food”: the two farthest halls accounted for 78 percent of refunds, but median kitchen-to-door time in the near halls was only 38 minutes. The root cause was the kitchen closing at 8:15pm combined with a 35-minute thermal window, not courier speed.

  • Theme one: cold food concentrated in two halls (evidence: 321 of 412 refunds).
  • Theme two: batching twelve orders starved the far routes (evidence: dispatcher logs plus courier exits).
  • Theme three: per-drop pay punished long trips (evidence: 14 of 19 exits cited pay).
  • Theme four: the 35-minute promise was never validated in winter (evidence: no baseline test).

Two things make this output trustworthy. First, every theme carries a count, so a claim like “couriers were slow” has to survive contact with a number, and here it did not. Second, the “what we controlled” split keeps weather and exam week out of the action list, which stops the retro from assigning a task to something nobody can change. If a returned theme has no supporting evidence, delete it before the meeting rather than letting it set the agenda.

The three actions that came back were: close the distant halls’ orders from a dedicated pickup point, pay a distance bonus on routes over 2.5 kilometers, and re-baseline the promise with a two-week test before reinstating it. Each had an owner and a metric. That is the shape you are aiming for, and it maps cleanly onto the plan we describe in our AI project plan workflow.

Stretching the Prompt to Other Runs

The same block adapts to almost any review with three small edits.

  • Post-mortem after an incident. Change “quarterly” to “after an outage”, keep the root-cause section, and tighten the action window to 7 days.
  • Monthly team retro. Cut the sections to themes, what worked, and actions, and run it on the last Friday of each month, as covered in our AI weekly review workflow.
  • Launch review. Add a section for “signals we ignored” and require one counter-metric per action.
  • Cross-team retro. Add an owner-name column so disagreements surface as gaps rather than arguments.

If you would rather keep the plan and the retro in one loop, fold the three actions straight into next quarter’s weekly plan using the structure in our weekly work plan prompt.

Run the retro within two weeks of the quarter ending. Memories fade fast, and Trayline’s drivers who quit in March could not be asked what went wrong by the time the meeting was finally scheduled in May, which removed the clearest evidence from the room. A retro held while the people who lived it are still reachable is worth more than a tidier one held later.

Sources & Notes

The section order here, evidence, themes, root causes, actions, is our own framing of standard retrospective practice, reviewed on 2026-10-12. Trayline, Priya Raman, Marco Vidal, and the north residence hall pilot are illustrative examples, and the metrics are representative rather than actual. Language models change between versions, so re-check the prompt and confirm any figure you carry into a real team meeting before you act on it.

声明:1.本站所有文章,如无特殊说明或标注,均为本站原创发布。本网站中【文本生成/文生图】使用了DeeSseek人工智能生成内容技术, 算法来源说明:本网站未自行开发、训练或部署深度合成算法,所使用的人工智能算法来自:DeepSeek、Chatgpt、Google Gemini。 2.为了网站加载速度优化,本站预览图片进行压缩处理,所以有些不是太清晰,并不影响AI提示词生图、AI设计等内容。3.任何个人或组织,在未征得本站同意时,禁止复制、盗用、采集、发布本站内容到任何网站、书籍等各类媒体平台。如若本站内容侵犯了原著者的合法权益,可联系我们进行处理。--GUUNN.COM