The filing happens without you. Put it on a schedule and it sweeps your inbox, distils your calls, reads your calendar and email, reconciles your to-do list and hands you the day. The one thing it cannot get anywhere else is why you decided something. That part takes a sentence.
Client and contact names blurred. Everything else is exactly as it sits, five months in.
Click to enlarge01 The field that pays you back. Not what happened, what was chosen and why.Click to enlarge02 One line per project, rewritten on every capture, with every entry and action hanging off it.Click to enlarge03 The to-do list, grouped by project and built from the record rather than typed.Click to enlarge04 The phone inbox. Ticked once the scheduled sweep has filed each one properly.Click to enlarge05 Capture between jobs. One line, no structure, no thinking about where it goes.
01 · The problem
Everyone remembers what they did. Almost nobody remembers why they chose it.
Project admin fails the same way every time. Capture costs more than it returns, so you stop doing it. Then the weekly review has nothing to review, and the to-do list drifts out of date until you no longer trust it.
The fix is not more discipline. It is removing the cost at the point of capture, which is the only place it matters. You talk for fifteen seconds. The system does the filing, the linking and the list.
02 · The automated day
Three scheduled runs, and the odd sentence from you
Once the schedule is set, the log maintains itself off what you already produce. Below is a normal weekday of mine. Note how little of it is me.
06:30Weekdays · scheduled
Morning sweep
Files every note dictated into the Notion Inbox since yesterday
Distils any new call transcripts, pulling commitments from both sides
Reads the calendar and triages what is genuinely waiting on a reply
Closes actions the log proves are done, flags the ones that look obsolete
Hands back the shape of the day and three to five tasks that fit the gaps
Your input: none
Ad hocYou, occasionally
The bit you actually do
Something happened that left no trace anywhere: a corridor conversation, a decision made in your own head, a call on your mobile
You say one line, or dictate it into the Inbox widget between jobs
The next scheduled run picks it up and files it properly
Your input: about fifteen seconds, when there is something only you know
The close-out and the weekly reviewTwo more scheduled runs
17:30Weekdays · scheduled
Close-out
Sweeps in anything logged since the morning
Summarises the day by project, a few lines, not a transcript
Flags what you mentioned but never turned into a task
Names any project whose current state now reads out of date
Sets tomorrow's top three
Your input: none
FridayWeekly · scheduled
The weekly review
Rolls up by client, then project, with each current state
Flags every active project untouched for seven days
Lists everything overdue, and everything sitting with someone else
Proposes next week's focus and waits for you to confirm it
Your input: one decision, at the end
Why it is on a timer, and how to run it without oneWorth reading once
Why scheduling is the whole thing. Every system like this dies the same way. Capture slips for three days, the review has nothing to review, and you stop trusting the list. Putting it on a timer removes the only step that depended on you remembering. The pack ships all three prompts, ready to paste.
You can absolutely do this by hand. Every routine above also works as a sentence. "Process my inbox." "Close out my day." "Weekly review." Nothing here needs a schedule to function, and plenty of people will run it that way quite happily.
Putting it on a timer is a Claude Pro feature, and what it buys you is not capability, it is consistency. The runs happen on the days you would have skipped them, which over a year is most of the value. Start manual. Add the timer when you already know you want it.
03 · How it works
When you do speak, this is what happens
The scheduled runs handle everything with a trace. This is the other case: something only you know, said badly, once. Left is what you say. Right is what lands in Notion.
What you say14 sec
just came off the acme call, they want
the pilot pushed to september. agreed
we cut scope rather than move the date
twice, because moving it twice kills
their trust in us. need to redo the SOW
by wednesday and i'm waiting on sarah
for the revised headcount
What lands in NotionOne write
Log entry
2026-08-10 · Acme Pilot · scope cut agreed, SOW to redo
CallAcme Pilot
What happened
Client call. Acme requested the pilot move to September.
Decisions · the field that pays you back
Cut scope rather than move the date a second time. Reasoning: a second date change damages client confidence more than a smaller pilot does.
To-do list, extracted
Redo the statement of work Due Wed
Chase Sarah for revised headcount Waiting
Project state, rewritten
Pilot moved to September with reduced scope; SOW being reissued.
Watch what happened to "waiting on sarah". It did not become a task you own. It became a Waiting item, which surfaces in the weekly review so you actually chase it. The commitments that go quiet are almost always the ones sitting with someone else, and most systems have nowhere to put them.
04 · What it reads
One record. Your list falls out of it.
Talking to it is the entry point, not the requirement. Connect the rest and most of the day files itself from things you already produce.
Inputs In the pack
01
What you say
Typed or dictated, in chat or into the Notion Inbox on your phone. Unstructured, unedited, mid-sentence.
Inputs Add a connector
02
Your calendar
Today's meetings and the real gaps between them. Google or Outlook.
03
Your email inbox
Triaged, not just listed. What is genuinely waiting on you, and what landed since yesterday.
04
Your meeting transcripts
Auto-transcribed calls distilled, not pasted. Commitments pulled from both sides of the call.
The work log
Five linked databases
Outputs
→
The log
What happened, what you decided and why, what is still unresolved.
→
Your to-do list
Built from the record, not typed separately. Every task linked to its source.
→
The daily plan
Three to five tasks that fit the time your calendar actually leaves you.
→
The weekly review
By client. Stale projects flagged. Overdue listed. Everything you are waiting on.
Nothing is filed twice. The to-do list is a by-product of the record, so it cannot drift out of date the way a separate task app does. It also closes actions only where the log proves they are done, never on the due date alone. Overdue is a flag, not evidence.
It also reads whatever else you connectBeyond the pack
Beyond the pack
It reads whatever else you connect
The five databases are the spine. Anything Claude can reach becomes another input, and the daily brief absorbs it. This is what mine reads on top of the core.
Analytics
Site traffic and enquiry volume, read each morning, so a quiet week is something you notice on Tuesday rather than in a month.
Scheduled tasks
Runs on a timer without being asked. Inbox swept, brief rendered, review prompted, whether or not you remember.
Team chat
Slack or Teams. Mentions and DMs still open on you, surfaced next to everything else instead of in a separate tab.
Anything else
Your CRM, your job software, your docs. If there is a connector, it becomes an input. Nothing in the schema has to change.
To be clear about what you are downloading: the pack contains the work-log skill and the Notion structure. The extras above are separate connectors you enable in Claude, and a second skill for the daily brief. Setup is per connector and takes minutes, but it is not one click.
05 · What it does
Four jobs, one set of data
01
Capture
Writes the entry, extracts the actions, links everything, rewrites where the project now stands.
02
Plan my day
Reads the calendar, sees the real gaps, hands you a shortlist that fits. Not a backlog.
03
Close out
Summarises the day, flags what you mentioned but never turned into a task, sets tomorrow's three.
04
Weekly review
Rollup by client. Stale projects. Overdue actions. Everything sitting with someone else.
+
Sweeps the inbox
Phone captures get filed into proper entries and ticked off. Say "process my inbox" and walk away.
+
Distils transcripts
Reads the full call, writes a short entry, links back to the transcript. Never pastes the whole thing.
+
Reconciles honestly
Closes tasks only on evidence from the log. Flags obsolete ones instead of quietly marking them done.
06 · Requirements
Short list
01Notion.Any plan. The free one handles databases, relations and rollups fine for one person.
02Claude, with skills and the Notion connector.Desktop, Cowork or Claude Code.
03Claude Pro or above, to put it on a timer.Optional. Scheduling lives on the paid tiers. Every routine still runs on demand without it.
04A calendar connector.Optional. Google or Outlook. Powers "plan my day".
05An email connector.Optional. Turns the morning brief into a real triage rather than a task list.
06Call transcription into Notion.Optional. Everything else works without it.
No API keys. No scripts. No terminal on Desktop or Cowork. Claude Code users get one command, and only because they asked for it.
Can I run this on the free plan?Yes, with one trade-off
On plans. The system itself runs on Free. Skills are on every Claude tier, Notion sits in the connector directory which is open to all users, and Notion's own free plan is plenty for one person. You can build the whole thing and drive every routine by hand without paying anyone.
Two things change when you go paid. Scheduling becomes available, so the routines run without being asked. And the heavier sweeps stop bumping into a free allowance halfway through. Neither adds a feature you cannot reach manually. Run it by hand for a fortnight and you will know whether that is worth a subscription to you.
Settings, then Connectors, then Notion. In Claude Code, add the Notion MCP server and run /mcp to confirm. Make sure the Work Log page sits inside the shared scope.
Test it: ask Claude to search your Notion for a database called Next Actions. If it finds it, you are connected.
02Build the Notion databases10 min
Create an empty Notion page called Work Log. Then paste the prompt below and Claude builds all five databases to spec. Or skip the prompt entirely: duplicate the ready-made Notion template, which is the more reliable route if you would rather not check an LLM's work.
Paste into Claude
Build my work-log system inside my Notion page called "Work Log".
Create five databases as full pages inside it, exactly to this spec.
Clients - Name (title), Type (select: Client, Internal), Status
(select: Active, Dormant, Archived), Primary contact (text), Context (text).
Projects - Name (title), Client (relation to Clients), Status (select:
Active, On hold, Blocked, Done), Type (select: Client work, Own product,
Admin), Priority (select: High, Medium, Low), Current state (text),
Stack / context (text), Last activity (date).
Log - Entry (title), Date (date with time), Project (relation to
Projects), Type (select: Dev, Meeting, Call, Planning, Research, Admin),
What happened (text), Decisions (text), Blockers / open questions (text).
Next Actions - Action (title), Project (relation to Projects), Source
log (relation to Log), Status (select: To do, Doing, Waiting, Done,
Cancelled), Priority (select: High, Medium, Low), Due (date), Notes (text).
Inbox - Note (title), Captured (created time), Processed (checkbox).
All text fields are Notion "Rich text". The Log "Date" must have Include
time switched on.
Create the relations in both directions and name the reverse sides exactly:
Clients "Projects", Projects "Log entries", Projects "Next actions",
Log "Spawned actions".
On Projects, add a rollup called "Open actions": relation Next actions,
property Status, calculate Count, filtered to Status is not Done and is
not Cancelled.
Then create one Client called "Internal" with Type = Internal, Status =
Active. List every property you created so I can check it. Tell me when
it is done.
Duplicate the Notion template · or build by hand from the schema. Whichever route you take, check the result against the schema. Notion names reverse relations its own way, so rename any that do not match.
03Install the skill2 min
Upload work-log.skill where your Claude surface accepts skill files, or paste SKILL.md into a new skill.
Claude Code
cp -r skill/work-log ~/.claude/skills/work-log
Restart Claude afterwards. Skills load at startup.
04Add two views3 min
Optional, but the system is far better with them. In Next Actions: "Today and overdue". In Projects: "Stale", filtered to Status is Active and Last activity before one week ago.
That second view is the one that catches work quietly slipping.
05Prove it works1 min
Say this. You should get a log entry, one task, and a project created or matched.
Test capture
Log this. Spent an hour setting up my work log system. Decided to
use Notion rather than a spreadsheet because I want the relational
links. Need to test the daily close-out tomorrow.
06Put it on a schedule10 min
This is the step that turns it from a tool into a system. In Cowork: sidebar, Scheduled, New task, Set up manually. Paste the prompt, set the frequency, save.
Where the output lands. Each run appears as its own session in your Scheduled list, and you open it to read the brief. Turn on notifications for the task if you would rather be told it has finished than remember to look.
Check the clock. Set the time in the scheduler and confirm which timezone it uses before you trust a 06:30. An hour out either way makes a morning brief useless.
Which surface. Scheduling lives in Cowork. Claude Code has its own separate scheduling, set up differently. If you are on Desktop without either, run the three prompts by hand and you lose nothing but the timer.
The pack contains all three prompts in SCHEDULING.md. Start with the morning sweep alone. Add the other two once you trust what it files.
Scheduled task · weekdays 06:30
Run my morning work-log sweep. Use the work-log skill. Work through
these in order and do not stop to ask me anything.
1. Process my Notion quick-capture Inbox. File every unprocessed note
into a Log entry against the right project, extract any next actions,
and tick Processed. If a note is too vague to place confidently, leave
it unprocessed and list it for me at the end.
2. Process any call transcripts I have not logged yet. Distil each into a
Log entry, pull out the commitments on both sides, link back to the
transcript, and mark it logged.
3. Reconcile my open Next Actions against the last 10 days of Log
entries. Close only those the log clearly proves are done. Never close
on the due date alone. Flag anything obsolete rather than closing it.
4. Read today's calendar and my email. Identify what is genuinely waiting
on a reply from me.
5. Give me a short brief: the shape of today, three to five actions that
realistically fit the gaps between my meetings, anything overdue,
anything I am waiting on someone else for, and anything you could not
file. Keep it tight.
One thing that catches people out: every scheduled run is a fresh session. It remembers nothing from your chats or its own last run. That is why the prompt repeats context you would never type in conversation. Keep that habit when you write your own.
08 · Free setup help
I will set it up with you. No charge, no pitch.
Twenty minutes on a call. We build the databases, connect it, and run your first real capture together so you leave with something working rather than something downloaded.
There is no upsell at the end. I am giving this away because I want to see what people do with it, and because the variants people build are more interesting than mine.
Once a month, on automation and VAs, and why they work. Every issue comes with a new free resource. You never need the list to get them. Everything here stays free and ungated. The newsletter is just how you hear about the next one.
09 · Reference
The schema
Five databases in one parent page. Property names are load-bearing: the skill searches by name, so rename in both places or not at all.
Open the schemaFive databases, every property
Create one entry immediately called Internal, Type set to Internal. Your own products and your own admin point at it. Without it, personal projects have nowhere to live.
Property
Type
Options
Name
Title
Type
Select
Client, Internal
Status
Select
Active, Dormant, Archived
Primary contact
Rich text
Context
Rich text
Projects
Relation
to Projects. Reverse side of the Client relation, name it exactly this.
These names are yours to change. Type and Stack / context are what suit my work; rename them to whatever your week actually looks like. Change them in the skill file too and nothing breaks. Current state earns its keep. One line, rewritten every time you log. It is what you read coming back to a project after two weeks, when you cannot remember where you left it.
Property
Type
Options
Name
Title
Client
Relation
to Clients
Status
Select
Active, On hold, Blocked, Done
Type
Select
Client work, Own product, Admin
Priority
Select
High, Medium, Low
Current state
Rich text
One line, kept fresh
Stack / context
Rich text
Last activity
Date
Log entries
Relation
to Log. Name the reverse side exactly this.
Next actions
Relation
to Next Actions. Name the reverse side exactly this.
Open actions
Rollup
Next actions → Status → Count, filtered to not Done and not Cancelled
Decisions is the highest-value field in the system. Not what you did, but what you chose and why. Six months out it is the only part you actually need. Titles follow YYYY-MM-DD · Project · short description.
Property
Type
Options
Entry
Title
Date
Date
Include time
Project
Relation
to Projects
Type
Select
Dev, Meeting, Call, Planning, Research, Admin
What happened
Rich text
Decisions
Rich text
The choice and the reasoning
Blockers / open questions
Rich text
Phrased as questions
Spawned actions
Relation
to Next Actions. Name the reverse side exactly this.
This is your to-do list, and you never type into it directly. Waiting is not decoration: it is for work sitting with someone else, and the weekly review surfaces it so you chase it.
Property
Type
Options
Action
Title
Verb first, physically actionable
Project
Relation
to Projects
Source log
Relation
to Log
Status
Select
To do, Doing, Waiting, Done, Cancelled
Priority
Select
High, Medium, Low
Due
Date
Notes
Rich text
Deliberately minimal. Three fields, so capture from a phone takes seconds. Put the Notion widget on your home screen and dictate a one-liner. Name the project inline if you can.
Property
Type
Options
Note
Title
The raw one-liner
Captured
Created time
Automatic
Processed
Checkbox
Ticked by the skill
10 · Daily use
The prompts worth knowing
You do not need to memorise commands. Just talk. But these are the ones you will use every day.
The prompts worth knowingEverything you will actually say
Say this
You get
Just describe what you did
A filed entry, tasks extracted, project state rewritten
Plan my day
Three to five tasks that fit your real calendar gaps
Close out my day
Today summarised, loose ends flagged, tomorrow's top three
Process my inbox
Every phone capture filed and ticked off
Process my call transcripts
Meetings distilled, commitments pulled from both sides
Weekly review
By client. Stale, overdue, waiting-on.
What am I waiting on?
Everything sitting with someone else, ready to chase
The one habit that makes this work: say the why, not just the what. "Moved the deadline" is worth half as much as "moved the deadline because their finance team signs off on the 5th and we were always going to miss it." The second one is the note you thank yourself for.
11 · Before you ask
Questions, and what goes wrong
The skill does not trigger when I describe my work
Say log this explicitly for the first few captures. A direct instruction always works.
Check the skill is installed and enabled in your skills list.
Restart Claude. Skills load at startup.
Claude cannot find my databases
Almost always connector scope. The Notion connector only sees pages explicitly shared with it. Open your Work Log page and confirm the connection is granted there or at a parent above it.
Test with: "Search my Notion for a database called Next Actions."
Entries file to the wrong project
Name the project in your dump. "Acme migration, got the import running" removes all ambiguity.
Merge duplicate projects. Two similar names is the usual root cause.
Keep names distinct. "Acme" and "Acme Phase 2" will get confused. "Acme Migration" and "Acme Reporting" will not.
Does it work without a calendar or transcripts?
Yes. Capture, the to-do list, close-out and the weekly review all run on your voice input alone. Calendar powers "plan my day". Transcripts remove the need to dictate after calls. Both are additive.
Do I need a paid plan?
Not to start. Skills are available on every Claude tier including Free, and Notion is a directory connector, which is open to all users. Notion's free plan handles the databases, relations and rollups without complaint.
Not to run it. Every routine works on demand with a sentence, on a free account. Paid tiers add scheduling, so it happens without you asking, and more headroom for the bigger sweeps. Try it by hand first and decide from there.
Where does my data actually go?
Your records live in your Notion, in your account. Nothing is stored by me and there is no server of mine anywhere in this. Claude reads and writes to Notion through the connector you authorise, and you can revoke that at any time.
Worth being straight about the rest: whatever you connect, Claude processes. That includes email contents and call transcripts if you connect those. Check Anthropic's and Notion's current data policies yourself rather than taking my word, and check your own client agreements before you point it at transcripts of client calls. Some contracts have views on that.
If in doubt, run it with the Notion databases only. The core system never needs your email.
Can it break or delete anything?
It creates and updates. It is not asked to delete, and the prompts tell it to flag rather than remove anything that looks obsolete. The realistic failure is a misfiled entry or a project state rewritten less well than you would have, both of which you can just edit.
Notion keeps page history, so an overwrite is recoverable. If you would rather check before it writes, set the scheduled tasks to require approval for the first fortnight.
Can someone else use it with me?
The structure is fine for it. Share the parent page and a colleague can capture into the same log, and the weekly review then covers both of you. Notion's free plan limits guests, so a second regular user is where you would start paying Notion rather than Claude. It was built single-user, so treat two-person use as something to test rather than assume.
I already use a task app. Does this replace it?
It replaces the task app, not the software you run your business on. Next Actions is a to-do list that happens to be generated rather than typed, so running it alongside Todoist or Asana means maintaining two lists and trusting neither.
Your CRM, job software or project tracker stays exactly where it is. This sits on top and reads from it if there is a connector.
How long before it is actually useful?
Week one is setup and getting into the habit. The first real payoff is the first weekly review, because that is the first time it tells you something you did not already know. The Decisions field takes a few months to start earning, and then it does not stop.
What actually happens on a scheduled run?
It opens a fresh session, works through the prompt, writes to Notion and reports back. It runs remotely, so your machine can be shut.
Two consequences worth knowing. It remembers nothing between runs, so the prompt has to stand alone. And nobody is there to answer questions, so the prompts tell it to leave anything ambiguous unfiled and list it for you rather than guess.
If you would rather check its work before it writes, set the task to require approval for the first couple of weeks.
Will it invent work I did not do?
The prompts are written to stop it padding, and in daily use it does not. If you capture nothing and take no calls, the close-out has nothing to summarise and says so. Read the first week's output properly anyway. Automation removes the admin, not the doing.
Can I use it with a different AI tool?
The Notion structure is tool-agnostic and worth building on its own. The skill is plain markdown describing a workflow, so it adapts with modest editing. Nothing is locked.
What are the honest limitations?
It only knows what you tell it. Garbage in, filed garbage out.
It will sometimes ask which project you meant. That is deliberate. A misfiled entry costs more than a five-second question.
The weekly review is only as good as the week of capture behind it.
Connector rate limits mean a large inbox sweep takes a minute. Sweep daily, not weekly.
Can I change it?
Please do. Fork it, rename it, strip out what you do not use. If it saves you an hour a week, tell me what you changed. The variants are the interesting part.
12 · Download
Take the whole thing
The skill file, the three scheduled-task prompts, the setup guide, the schema spec, the prompt sheet and the troubleshooting notes. No email needed to download it, and no sequence afterwards.
Eight files · 38 KB unzipped · Claude Desktop, Cowork and Claude Code
Would rather not download anything?
Paste it into a new skill in Claude. You will still want the setup steps above for the Notion side.
13 · What it unlocks
Once the log is live, other things fall out of it
The record is the asset. Both of these run on the same five databases and the same scheduled sweep, with no extra input from you.
Coming next
The daily brief, rendered
The morning sweep already gathers everything. This turns it into a single page you actually want to look at: the shape of the day drawn out, what needs you, what is already handled. Its own resource, because it deserves its own setup.
LinkedIn content, generated from your live work log
You already record what you did and why you chose it. That is the raw material for posts most people struggle to find. A second skill turns real entries into drafts in your voice, without inventing anything you did not do.
In progress
14 · Who made it
Alex Davis
Founder · Operations Oversight Advisory
alex@oversightadvisory.io
I built this because I was losing the reasoning behind my own decisions across half a dozen projects, and no amount of discipline fixed it. The problem was never discipline. Capture cost more than it returned, so I stopped doing it.
I have used it every working day since. It is the same approach I take into owner-led businesses: find the point where the process actually breaks, and take the cost out of it, rather than adding another rule on top.
