how to write your weekly update from your own work

how to write your weekly update is not really a writing question. it is a recall question, and every template on the internet solves the wrong half.

you sit down on friday, open the doc, and cannot remember monday. so you scroll back through slack, skim your commits, check your calendar, and reconstruct an approximation. the reconstruction takes an hour and the writing takes 10 minutes.

here is a template that works, and then the part underneath it that no template fixes, current as of september 2026.

the template

a good weekly status report has 3 sections and fits on one screen.

the whole discipline sits in column 3 of row 1. write the effect, not the activity. "worked on checkout" tells your reader nothing. "checkout conversion recovered after the retry fix shipped tuesday" tells them what is different about the company this week.

a weekly update template that asks for more than these 3 sections will get filled in badly, because the marginal section is always the one people pad.

write for a reader with no context

the audience is someone who was not in the room, and that changes every sentence.

your manager has 8 of these to read. an investor has 20. neither remembers the acronym you used last month or which project the ticket number belongs to. so name things in full the first time, say why a result matters rather than assuming it is obvious, and lead each bullet with the outcome rather than burying it after the setup.

this is also the difference from a standup update. a standup is for people already holding the context, covers 1 day, and exists for coordination. a weekly update is a briefing for someone who has been elsewhere, and it is the one that gets quoted back to you in a performance review 6 months later.

the reconstruction problem

you cannot remember the week because you never encoded it, and that is normal rather than a personal failing.

attention during work is on the problem, not on recording that you worked on it. by friday the trace is weak. worse, the parts you remember best are the ones with emotional weight, which are usually the frustrating ones rather than the important ones, so unaided recall systematically over reports what annoyed you.

there is a second effect stacked on top. people who know information is stored somewhere retrievable remember the location rather than the content, an effect documented by betsy sparrow and colleagues in science in 2011. you remember that the decision is in a thread somewhere, not what it was. the fuller mechanism is in why you forget what you read.

the artefacts you can actually pull

your tools hold fragments, and the work is in the join.

each row is 10 to 15 minutes of scrolling, and none of them covers reading, thinking, or the 3 days you spent on an approach you abandoned. an honest work log built only from these sources over reports shipping and under reports everything that made shipping possible.

what the usage data says about this

people already spend more ai time searching than doing anything else, and almost none of it reaches their own week.

per the anthropic economic index for may 2026, searching electronic sources is the most common work task in sampled claude conversations at 4.95%, with reference searching second at 3.74%.

content creation is 22.72% of conversations, which is where drafting an update lands. but knowledge retrieval and enterprise search is 3.61% and conversation and meeting intelligence is 0.26%, and work is 43.36% of classified conversations overall. people are drafting far more than they are recalling, because there is nothing to recall from.

that imbalance is exactly why ai drafted updates read as generic. the model has the writing ability and none of the week.

can ai write it for you?

it can write a good first pass from evidence, and it cannot invent the evidence.

hand an assistant your commits, closed tickets and calendar and it will group them sensibly and write clean prose in seconds. claude's memory and the equivalents in chatgpt and gemini help with tone and preferences, and none of them holds your week. hand it nothing and it will produce confident filler that sounds like every other update, which is worse than a rough honest one.

so the useful question is not which model. it is what the model can read. an assistant connected to your own history can answer "what did i do this week" from the record instead of from your memory, and connecting one is a config edit rather than a project. the mechanics are in what is an mcp server, and the wider case is how to give an ai agent context about your own work.

a friday routine that takes 15 minutes

5 steps, in order, and none of them require remembering anything.

1. pull the artefacts. commits, closed tickets, sent messages, calendar. 5 minutes.
2. search your own history for the week. this is the step that catches the reading and the abandoned approaches.
3. group by outcome, not by day. nobody wants a diary. they want 3 things that are now true.
4. write the blocked section honestly. an update with no blockers is either a very good week or a hidden problem.
5. cut activity that produced nothing. it feels like padding because it is.

step 2 is the one most people cannot do, and it is the difference between a reconstructed update and an accurate one.

where remynd sits

remynd records the focused window on a mac, runs ocr locally through apple's vision framework so what you looked at becomes searchable text, and keeps the index on your machine. that is what makes "what did i work on tuesday afternoon" a question with an answer.

the scope, stated exactly: focused window capture rather than every pixel of every display, local ocr and storage, recordings default to 30 days of retention, app and site exclusions, and history that is read only everywhere in the app. call transcription runs on device using an mlx speech model on apple silicon. in august 2026 it shipped a local connector, so claude code and codex can query that history when drafting.

what it does not do: asking a question sends the retrieved slices to a cloud model, so this is not a claim that nothing ever leaves the machine.

if you want the category rather than the routine, start with screen memory.