My Project Plans Were Two Months Out of Date. I Rebuilt Them From the Record.
Situation
I built my project plans during my second week at a new job.
I'd sat through a stack of knowledge transfer meetings and understood maybe a third of what I'd heard. I needed plans anyway. I had AI help me put them together from what I had, and it filled the gaps the way it does, adding tasks that sounded reasonable for a program like this one. Some of them probably were. I had no way to tell which.
Two months later I opened one of those plans to pull it into a working meeting and read it top to bottom.
Nothing in it was wrong, exactly. It just didn't mean anything to me. Tasks I couldn't connect to anything real. Phases that didn't match how the work was actually moving.
And I was about to walk teams through it. If it didn't make sense to me, it wasn't going to make sense to them. I was also building a walking deck off the plan, something I'd be maintaining every week, so whatever was in the plan was about to become the thing everybody looked at.
The AI Move
I'd been in every one of those meetings. I just couldn't hold months of them in my head.
The record could. I pointed Claude at Zoom, Slack, Drive and Gmail and told it to find anything and everything about the program, starting from my first day. About fifteen meetings with transcripts. One Slack channel going back four months. A handful of emails. One Drive folder.
Then I asked for two things.
Rebuild the history: every task that had already been done, and when it actually finished, pulled from wherever it was mentioned.
Then build the forward plan from what people had actually said over those four months, not from what a plan like this usually contains (meaning an AI-generated plan).
For the plan itself I used the Smartsheet connector, with one rule. Don't update anything on your own. Add the new rows at the bottom and leave everything already there alone.
I wanted to read the new plan sitting next to the old one before either replaced anything. So that's what I did. New rows at the bottom, my rows at the top, going line by line looking for anything of mine that should have made it into the new plan but didn't. When I was satisfied nothing important was lost, I deleted the old rows myself.
The Shift
I expected to argue with it.
I read through the reconstructed history and kept nodding. Yes, that happened. Yes, that's when it finished. It surfaced work from month one that had closed out and left my head completely. In the forward plan there were tasks I'd never written down anywhere, things agreed in a meeting or settled in a Slack thread that never made it onto the sheet.
I had very few questions by the end, which is not how these usually go for me.
What changed wasn't the quality of the plan so much as where it came from. I wasn't reconstructing anything. Two months of my recall against two months of what was actually said and written, and the record won every time they disagreed.
The Pattern
The plan I built in week two did what a plan does at that stage. It gave me something to point at while I learned.
Where I went wrong was leaving it there. I kept updating it, when what it needed was to be rebuilt once I knew enough to know what belonged in it.
I don't think I'll have AI build one from scratch again. That's where the noise comes from. Ask for a plan with thin context and you get a complete one back, because complete is what you asked for, and you can't tell the real tasks from the plausible ones. Building as I go and having it check my work against the record gets me a plan where nothing is in there unless it came from somewhere.
The Implication
Open a plan you wrote early in something and read it as though someone else handed it to you. Ask whether the tasks in it still describe real work.
If they don't, don't fix it from memory. You've forgotten more than you think you have. Point AI at wherever the actual conversations live, the meeting transcripts and the channel where the work gets discussed, and ask it to rebuild from there.
Then compare instead of replacing. Put the new version next to the old one and delete nothing until you've read both.
What I'm Testing Next
The rebuild was the one-time part. What I set up afterward is a scheduled task that runs once a week and reconciles the plan against whatever surfaced in the last seven days.
It flags delays, dependencies nobody had written down, and tasks that got completed and mentioned in Slack but never came up in a meeting. That last one happens more than I would have guessed. It proposes, I approve, and nothing moves in the sheet without me.
I'm rolling that into every program I run from here. Building the plan is the part I'll keep doing myself. Keeping it honest is the part I'm handing over.
A Note on the Connectors
The only thing that makes this work is AI being able to read the actual record instead of you pasting pieces of it into a chat window. That's what the connectors do. Zoom for the meeting transcripts, Slack for the channel history, Drive and Gmail for whatever else got written down, and Smartsheet for the plan itself so the updates land where the plan already lives.
If you don't have those, this still works with more typing. Export the transcripts, copy the channel history, and do the sheet updates by hand at the end. The rebuild is the valuable part. The writing back is just typing.
The Plan Reconstruction Prompt
Run this in stages and stop for approval between each one. Do not let it write to your plan until Stage 5.
Stage 1: Read the record. Go through my meeting transcripts, my team's channel, my email, and my project folder, starting from [date]. Tell me what you found before you interpret any of it: how many meetings, what date range, which recurring meetings exist, who the recurring participants are, and where the gaps are. If a stretch of time has no coverage, say so now rather than filling it in later.
Stage 2: Rebuild the history. List every task, decision, or deliverable that has already been completed. For each one give me the task, the date it was completed, and where you found it, quoting the line that told you. If you can only establish a rough date, say so and give me the range instead of picking a date. Do not include anything you cannot point to a source for.
Stage 3: Build the forward plan. Now list the work that is still ahead, based only on what people have actually discussed. For each task give me the source, the owner if one was named, any date or dependency that was mentioned, and how confident you are that this is real committed work versus something raised once and never picked up again. Sort it so the low-confidence items are at the top where I will read them. Do not add standard tasks that a program like mine would normally have. If you think something obvious is missing, list it separately under a heading that says these are your suggestions, not findings.
Stage 4: Compare against what I already have. Here is my current plan. Show me what is in yours and not mine, what is in mine and not yours, and where the two disagree on dates or status. For anything in mine that you did not find in the record, do not assume it is wrong. Flag it and ask.
Stage 5: Write it as new rows only. Once I approve, add the new plan as rows at the bottom of my sheet. Do not edit, overwrite, or delete any existing row for any reason. I want to read both versions side by side and I will remove the old ones myself. When you are done, read the sheet back and show me what is actually there.
My program is [what it is]. I started on [date]. My plan lives in [tool]. Start with Stage 1.
Comments