Posts

Showing posts from October, 2026

Long chats go stale: write a restart note and begin again

Written by the Octuo team at Tutuo, Inc. Somewhere deep into a long conversation, the thread starts to behave oddly. The helper proposes an idea you rejected an hour ago. It answers a question you have since changed. It treats an early draft as the final one. Nothing is broken. The thread simply holds too many versions of the truth. You then face two poor options: keep going and keep correcting, or start from scratch and explain everything again. There is a third: write a restart note. Signs the thread has gone stale You are correcting the same misunderstanding for the second or third time The helper cites a decision you reversed Replies get longer but less specific You cannot remember what was decided without scrolling You have pasted the same background twice If two of these sound familiar, the thread has done its job and it is time to move on. The restart note: six lines Goal. One sentence, in its current version, not the version you began with. Decisions in force. Only the ones tha...

Set a retry budget: three attempts, each changing one thing

Written by the Octuo team at Tutuo, Inc. The task came back wrong. You ask again, a little more firmly. It comes back wrong in a different way. You ask once more. Forty minutes and six runs later, you have learned very little. Retrying out of habit feels like progress and is usually just spending. Every run costs something: time, attention, and often money. A retry budget is a rule you set before the first run, so that the sixth attempt never happens by accident. Why decide in advance Mid-task, stopping is hard. Each failed attempt makes the next one feel more justified, because you have already put so much in. That is the sunk-cost trap, and it is nearly impossible to resist in the moment. A number you chose while calm does the resisting for you. It also changes how you work: if you only have three attempts, you spend the first one thinking about what to change. The rule Set the number before you start. Three is a sensible default. Use fewer for expensive runs and more for quick, chea...

Six places a secret hides after you have used it

Written by the Octuo team at Tutuo, Inc. You paste an API key into a terminal to test something. The test passes, the key still works, and you move on. Where is that key now? Most credential leaks are not dramatic. They are leftovers: a copy in a history file, a line in a log, a corner of a screenshot. This article lists six places to check after any session where a secret was in play, whether or not an assistant was involved. The six places Shell history. Commands typed with a key inline are usually saved in a history file that can outlast the task. Load keys from an environment file that lives outside your project, and delete the history line if one slipped in. Config and environment files inside a project folder. Project folders get copied, zipped, committed and shared. Keep secrets out of them, add the file name to your ignore list, and look before every share. Logs and error output. Failing tools sometimes print the URL, the headers or the whole configuration, token included. Read...

Turn the job that went well into a template: five things to replace

A job went well last month. The brief was clear, the result needed one small fix and you were pleased. Now it comes round again, and you are rewriting the brief from scratch, because the old one is full of last month's names, dates and file names. (A hypothetical scene, for illustration.) A job that went well is worth keeping in a form you can reuse. Turning its brief into a template takes about ten minutes, and it works for a job handed to a colleague, a contractor or an AI assistant. Five things to replace: Replace the specifics with labelled blanks. The month becomes MONTH, the person's name becomes NAME, the file becomes FOLDER. Capital letters tell the next reader what to fill in and what to leave alone. Keep the wording that worked, word for word. If one sentence fixed the format last time, leave it exactly as it is. Resist improving the parts of the brief that did their job. Add a "what changed last time" note. A line such as "Last month the totals row was...

Rough, usable or final? Say which finish you want before the work starts

You asked for a first look at a flyer layout and received something polished enough to print. You wanted a sketch to react to. The helper delivered a finished page, and now you feel awkward asking for big changes to something that looks done. (A hypothetical scene, for illustration.) People read "draft", "version" and "a quick look" in very different ways. One short word in the request settles it: say which finish you want. Four levels cover most jobs, and they work with any helper, whether a colleague, a designer or an AI assistant: Rough, for deciding direction. Quick, incomplete and even ugly is fine. You want to know whether the idea is right, and nobody should polish anything that may be thrown away. Say so: "Rough. I will react to the idea, not the details." Working, for trying it. It does the job from start to end, but the corners are not rounded. It is good enough to run once with real material and see what happens. Usable, for sharing wi...

Checking a summary against its source: five checks in ten minutes

A helper reads a forty-page report and hands you a one-page summary. It reads well. You are about to forward it to your manager when you realize you have no idea whether the report actually says any of it. (A hypothetical scene, for illustration.) A summary is a claim about a source, and claims can be checked. You do not have to read the whole report again. Five checks find most of the problems in about ten minutes, whoever wrote the summary: a colleague, an intern or an AI assistant. Pick three claims and find them in the source. Choose the ones the summary leans on most, and for each one find the page or paragraph it came from. A claim you cannot find is a problem you have found. Check every number against the source. Figures, dates and percentages get altered more often than any sentence. Look at units and periods too, because "per month" and "per year" slip in easily. Look for what is missing. Open the source's headings or contents and ask which section has ...

Before you paste an email thread into a brief: five things to leave out

You want a helper to draft a reply, so you select the whole email thread and paste it into the brief. It takes a second. It also carries along four signatures, a phone number, a candid remark about someone who has nothing to do with this job and a forwarded message from two years ago. (A hypothetical scene, for illustration.) A thread is often the best background a helper can get, but it was written for the people in it, not for a stranger who needs one answer from it. A two-minute trim before you paste keeps the useful part and leaves the rest behind. It works for any helper, whether a colleague, an agency or an AI assistant. Five things to leave out: Signatures and contact details. Phone numbers, home addresses and job titles rarely help with the task. If the helper needs a name, give the name. Remarks that were not meant to travel: a joke about a colleague, a comment on someone's performance, a half-formed opinion. If a remark matters to the task, restate the point in your own n...

Choosing what to keep human: a four-question filter

You made a list of everything you could hand off, and it came to nearly everything. That is a warning sign, not a plan. Some of that work is the reason you are in the role at all: the call you make when two people disagree, the sentence that says sorry, the decision nobody else can sign. (A hypothetical scene, for illustration.) Handing work off well starts with deciding what stays with you. A four-question filter takes about five minutes per task and works for any helper, whether a colleague, a contractor or an AI assistant. Ask the questions in order and note every yes: Would a mistake here be hard to undo? Sending, paying, publishing, deleting and signing are all hard to take back. A task that ends in one of them can still be prepared by a helper, but the last step stays with you. Does it carry your name in a way that matters to the reader? A thank-you, an apology, a difficult message or a reference. People can tell when a message was routine and when it was meant. Keep these, or at...

Silence is not success: how to notice a routine that did not run

Every Monday a summary lands in your inbox. This Monday it did not. You only notice on Thursday, when someone asks where it is, and you realize that last Monday's never arrived either. The routine did not fail loudly. It stopped, and a missing message looks exactly like a quiet week. (A hypothetical scene, for illustration.) A routine that runs on a schedule, such as a weekly report, a nightly backup or a Friday reminder, has an awkward property. When it works you see a result, and when it breaks you see nothing. Nothing is easy to miss. These five habits turn silence into a signal, and they work whether the routine is run by a colleague, a script or an AI assistant: Write down when it counts as late. "The summary arrives by Monday at 9:00, and at 10:00 it is overdue." Without a stated time nobody can say that a routine has missed, and everybody assumes somebody else has noticed. Ask for a result every time, even when there is nothing to report. "Ran Monday 8:02. No ...

From a rambling voice note to a clear brief: five edits

You talk through a task on the way to your next meeting: who it is for, what you want, the thing you forgot, the other thing you forgot and a deadline you only half recall. The voice note is a good record of what is in your head. It is a poor brief, because whoever receives it has to do the sorting you skipped. (A hypothetical scene, for illustration.) Talking first and editing second is a fine way to start. You get everything out before you worry about order. Five quick edits turn the rambling version into something a helper can act on, whether it is a colleague, a contractor or an AI assistant: Underline the outcome. Find the sentence that says what should exist when the job is done and move it to the top. Everything else supports that sentence. Separate the facts from the wishes. Dates, names, amounts and file names go in a short list of facts. "It would be nice if" goes on its own line, marked optional. Cut the second telling. When you said the same thing twice with diffe...

Following up without nagging: a short ladder for replies you are waiting on

You asked three people for their numbers by Thursday. It is now Friday morning and you have two of them. You could write "just checking in" to the third, which sounds like a nag and says nothing, or you could wait and hope. Most of us alternate between the two. (A hypothetical scene, for illustration.) Following up is a chore that is easy to hand to an assistant, a colleague or a tool, as long as the rules are written down first. A short ladder keeps it polite, specific and finite: Send the request with a date and a reason. "Could you send the June figures by Thursday at noon? I need them for Friday's review." A date and a reason make every later reminder easy and fair. Send the first reminder the day before the date, short and helpful. Restate the ask, attach whatever is needed and offer a way to make it easier: "Happy to take a rough number if the final one is not ready." Send the second reminder on the date itself, with the one specific thing and it...

Sort your requirements into must, should and nice before you ask for options

You ask for options for a team offsite and get twelve venues. None is wrong. But six are too far away, four are over budget and two have no room for lunch, and you knew all of that without ever saying it. (A hypothetical scene, for illustration.) Options come back sorted by whatever the helper guessed you cared about. You can steer them by sorting your own requirements first. The method takes ten minutes, works for any helper, and has five steps: Write every requirement on its own line, as it occurs to you, with no ranking yet. Include the obvious ones. What is obvious to you is invisible to everyone else. Mark each line must, should or nice. A must is a dealbreaker: an option that fails it is out. A should is something you would trade for a good reason. A nice is a tiebreaker and nothing more. Put a number or a test on every must. "Close to the office" is a wish. "Under forty minutes by public transport from the main office" is a test an option can pass or fail. Ke...

Before a helper edits your document, make a copy and ask for a change list

A colleague, or an assistant, tidies up your document. It reads more smoothly, which is nice. It also no longer says what you meant in paragraph four, a figure in the table has changed, and you cannot tell how many other things are different because you saved over the original. (This is a hypothetical scene, for illustration.) Editing is one of the most common handoffs and has the least safety net. Five habits give you one, and they work whoever does the editing: Copy first, then hand over the copy. Save the current version under a new name before anyone touches it, and give the helper the copy. The original stays exactly as it was, and every later question can be answered by comparing the two. Say what kind of editing you want. Proofreading, tightening, restructuring and rewriting are four different jobs. "Fix typos and obvious grammar only, and do not change wording or meaning" is a very different request from "make this clearer". Name the parts that are off limit...

Two drafts, one decision: a four-question scorecard for comparing them

You asked for two versions of an announcement, and now both are open side by side. Both are fine. You read the first, then the second, then the first again, and you pick the one you read last. That is not a decision. It is the order you looked in. (A hypothetical scene, for illustration.) Comparing drafts goes better with questions written down in advance, because they take the order of reading out of the result. This scorecard has four questions and takes about five minutes. It works for drafts from a colleague, a template or an AI assistant: Which one says what the reader should do, in the first two lines? Cover everything but the openings and read only those. The draft where the action is clearest has won the most important round. Which one could you send without checking a fact? Mark every claim, figure and name in each draft. The draft with fewer marks is quicker to finish, and a draft with a claim you cannot source is not finished at all. Which one sounds like you on an ordinary ...

Ask for a checkpoint halfway: what a good partial result looks like

You hand over a big job on Monday morning and hear nothing until Wednesday night. Then it arrives, finished, and it is a long way from what you meant. Two days of work went into the wrong shape, and all you can do is say so and start again. (A hypothetical scene, for illustration.) A checkpoint is a deliberate pause partway through a job: the helper stops, shows what it has so far and waits for a go-ahead or a change of direction. It works for any helper, a colleague, a contractor or an AI assistant, and it turns one big gamble into two smaller bets. A good partial result has five parts: A slice, not a sketch. Ask for one finished piece of the whole: the first three sections, the first fifty rows, one of the five pages. A complete small piece shows you real quality. An outline of everything only shows you intentions. The method in two sentences. How the helper is going about the job, and what it has assumed so far. Most wrong turns are visible in the method long before they show up in ...

Put a footer on every recurring report: five lines that keep it honest

A weekly report has a way of becoming furniture. It lands in the same inbox at the same time, people nod at the same chart, and nobody can say any more where the numbers come from or whether the person who set it up still owns it. The report keeps running long after it has stopped being checked. A short footer fixes most of that. It takes a minute to write, and it turns a report from something that appears into something that can be audited. It works for reports made by people, scripts or AI assistants, and it works best as a template that gets copied each time. Five lines: Sources. Name every file, sheet, dashboard or system the figures came from, and say which version or export you used. When a number looks wrong, this line tells you where to look first. Period. State the exact date range covered, with the time zone if it matters. "Last week" means different things on different days. Changes. Say what is different from the previous edition: a new source, a changed definitio...

Check in twice a day, not twelve times: set review windows for delegated work

You hand a job to an assistant and then you can't stop looking. Is it done? Did it get stuck? You open the window, glance, close it, and ten minutes later you do it again. The checking costs more than the job, and it chops your own work into pieces too small to finish. Review windows are a simple fix: you decide in advance when you will look at delegated work, and you look only then. For most people two windows a day are plenty, one in the late morning and one toward the end of the day, though the right times depend on when your results arrive. Here is how to set them up: Pick two or three fixed times and put them on your calendar like meetings. A window that is not on the calendar will not survive the first busy day. Decide what a job should do when it needs you in between. Say what counts as urgent enough to interrupt you (a decision it cannot make, a risk, a missing input) and where the interruption should go. Everything else waits for the next window. Ask for a one-glance statu...

Show one good example, and say exactly what to copy from it

You ask for "a short update in our usual style" and what comes back is cheerful, three paragraphs long and topped with a heading you never use. The helper did what you said. It just didn't know what "usual" meant, and you didn't know you hadn't said. The quickest fix is rarely a longer description. It is one good example. A single real sample carries dozens of small decisions that are tedious to put into words: length, order, tone, how numbers are formatted, where the date goes. But an example on its own invites copying too much, so a good handoff pairs it with a note on what to take from it. Here is a way to do that: Pick the best real example you have, not an average one. If your only example is a weak one, say so, and describe what you would change. Strip out anything private or one-off, such as names, amounts and dates, and replace each with a plain placeholder in square brackets. Label what to copy: the structure, the length, the order of points, th...

Four requests in one message? Number them so none gets dropped

You send one message with four requests tucked into it. The reply arrives quickly, is friendly and covers two of them. The other two were not refused. They simply disappeared into the paragraph, and a week later you are still waiting on something nobody remembers being asked. (A hypothetical scene, for illustration.) When several requests share a paragraph, the first and the last tend to get answered and the ones in the middle get lost. Numbering them takes about a minute and works with any helper, whether a colleague, an agency or an AI assistant. Five habits make a numbered message do its job: One request per number, and one action per request. "Confirm the room for the 14th" is one request. "Confirm the room and tell the caterer" is two. If you cannot give a request a number of its own, it is not finished yet. Put them in the order you need the answers, not the order you thought of them. If number three has to happen before number one, say so or move it up. Mark ...

Hand over access, not secrets: a five-step credentials habit

To get a task moving, it's tempting to paste a password into a chat, a token into a shared document or a login into a message. It works for an afternoon. Then the chat is forwarded, the document is copied and the password lives in more places than you can list. Delegating work to a person, a script or an AI assistant means giving it some kind of access. These five habits keep the access and lose the exposure: Keep secrets out of text. Anything pasted into a chat, a ticket or a document should be treated as copied. Share access through the system that issued it, or through a password manager, not through a message. One key per job. A reporting task gets a read-only key for reporting, not your main login. If the job changes, issue a new key instead of widening the old one. Name keys after what they do. "Weekly-report-readonly" tells you what to cut when the project ends. "My token" tells you nothing. Keep a short list. Service, purpose, date connected, and who or ...

Schedule a recurring job backwards from the moment you need the result

Most recurring jobs are scheduled from the wrong end. You pick a time that suits the job, and on Monday, five minutes before the meeting, you find that the summary you needed isn't ready, or is ready and unread. Work backwards instead. Five questions, in this order: When do I need the result, and where will I look for it? Name the moment ("before Monday's team meeting") and the place. When will I review it? A result you first read in the meeting hasn't been reviewed. Leave a gap long enough to read it and send something back. How long does the job take? Use the longest run you've seen, not the average, and add a margin. When do its inputs settle? A job that starts before the last input arrives reports on a version that is already out of date. Start after the inputs settle. When will I look at the schedule itself? Put a review date on it: "Check this routine at the end of next month." A recurring job nobody revisits keeps running long after anyone nee...

Going away for a week? Write the away note for your jobs, not just your inbox

Everyone writes the out-of-office message. Almost nobody writes the other one: what happens to the work that is still moving while you are gone. The recurring report that still runs on Friday, the half-finished research, the follow-up you promised on Wednesday. They don't know you left. A job away note is short, and it is written for whoever, or whatever, will be running things while you cannot answer questions. It works just as well if you are off sick or moving to another project. Five things belong in it: What keeps running, with its schedule and where the results land. List each recurring job and when it runs. What pauses, and what "paused" means. Say which jobs stop completely until you return, so that nobody wonders whether they were forgotten. What can be decided without you. Be concrete: "pick the lower-cost option of the two if the difference is under a set amount", "move a meeting within the same week", "send the reminder". Give the...

The four-line report: how to give feedback on any new tool so it gets better

You try something new, it does something odd, and you shrug and move on. A week later it does the same thing, and you still haven't said a word. The people who could fix it never heard about it. Useful feedback takes four lines, and it works for a tool in public beta or one that has been around for years. The example is hypothetical and illustrative: an assistant summarizes a long email thread and leaves out the one decision that mattered. What I asked. Paste the request as you actually wrote it, not as you remember writing it. What I expected. One sentence: "The venue decision should be in the first line." What happened. What came back, quoted, and what was missing or wrong. What would have helped. Describe a need, not a fix: "put decisions first" rather than "use a different model." Two small additions make a report much easier to act on. Say whether it happened once or every time. And if the surprise was about cost, include both numbers: what the es...

Five questions to ask before you click Allow on Screen Recording and Accessibility

Every so often an app asks for a permission that makes you pause. Screen Recording and Accessibility make people pause longest, and they should: one lets an app see what is on your screen, the other lets it operate other apps for you. Neither permission is good or bad in itself. What matters is whether you can say why the app needs it, what it can do with it and how you would take it back. Five questions help: What job needs this? Name the task before you grant the permission. "So it can work in my browser while I'm in meetings" is a reason. "It asked" is not. Which of the two do I need to grant? They do different things: Screen Recording is about seeing, Accessibility is about acting. If a job needs only one, ask why the app wants both. Do I know where the off switch is? Both live in System Settings, under Privacy & Security, and you can turn them off there at any time. Find the switch on the day you grant the permission, not the day you need it. Does the w...

How many results from a big batch do you actually have to read?

Say an assistant has just summarized forty meeting notes, or renamed four hundred files, or drafted fifty product descriptions. You can't read everything, and you don't want to trust everything. Somewhere between "read it all" and "looks fine" there is a sensible amount, and it is usually smaller than people fear. (A hypothetical scene, for illustration.) The idea is to review in layers: look at a little of everything cheaply, then look at a few of the risky ones carefully. This routine works whether the batch came from a person, a script or an AI assistant: Skim the whole batch for shape, not content. Are the items about the same length? Do any come back empty, or repeat word for word? Outliers in length and format are often where the first problems show up. Two minutes is enough. Pull a handful at random and read them in full. Choose by rule, not by feel: every seventh item, or every item whose number ends in 3. If you choose by eye you will pick the inter...

Run a three-item pilot before you hand over the whole job

Say you have a folder of receipt PDFs to rename and file by month. It's a good job to hand off, and a good job to get wrong at scale: if the first assumption is off, every file inherits it. (This is a hypothetical scene, for illustration.) A pilot is cheap and tells you a lot. The routine works for any repetitive job, whether a person, a script or an AI assistant does it: Pick three items on purpose: an easy one, a typical one and an awkward one. The awkward one is the file with the missing date, the odd layout or the strange name. Write the brief for the whole job, not for the pilot. The point is to test the brief. Run only those three. Read the results the way you would read the finished job. Check each one against your finish line. Did the awkward one get handled, or quietly skipped? Settle every ambiguity you noticed, in writing: which date to trust, what to do when a field is missing, where uncertain items go. Scale up in steps: a small batch, then a larger one, then the rest....

A model-per-task playbook for an ordinary work week

Choosing a model for each task sounds like more work than choosing one and sticking with it. It can be a short habit: ask what the task needs, then pick. Here is a playbook for five ordinary tasks. It names no specific model on purpose. Lineups change; the questions don't. Task 1: Triage, sorting a pile of notes or messages. What matters: speed and cost. Mistakes are cheap, because you'll look again. A lighter, faster option is often the natural fit. Task 2: First draft, a summary, a follow-up or an outline. What matters: tone and structure. You'll edit it anyway, so pick for the kind of writing you want and accept that it's a draft. Task 3: Long reading, a stack of notes or a long thread. What matters: staying accurate across a lot of material. This is where a stronger option may earn its cost. Spot-check two claims against the source before you trust the summary. Task 4: Anything that changes something, such as editing shared files or sending messages. What matters: c...

Name your files for the stranger who inherits them: five rules

You open a shared folder and find "final.docx", "final_v2.docx", "final_v2_REAL.docx" and "notes (3).txt". You know which one is current, because you were there when each was made. A colleague, a contractor or an assistant picking up the work has to guess, and a wrong guess quietly becomes the starting point for everything that follows. (This is a hypothetical scene, for illustration.) A file name is the smallest handoff note you will ever write. It is read before the file is opened, it travels with the file into every folder and email, and it is the one piece of context that nobody can detach. Five rules cover most of what goes wrong: Start with the date, year first. "2026-03-14" sorts correctly in every folder view, and it tells a stranger how old the file is without opening it. Name the project next, spelled the same way every time. If it is "q3-launch" in one file and "Q3 Launch Plan" in another, a search will fi...

How Octuo pricing works, in plain language

We'd rather you understood the pricing before you pay for anything, so here it is in plain language. Octuo is a personal assistant and one starting point for work across AI, tools and specialist services. It coordinates the supported tools and services you connect and authorize. Octuo runs on credits. Usage is measured in Octuo Credits (OC). One OC is roughly $0.001. Where credits come from. New accounts get 1,000 OC trial credits, once per person, no card. After you verify your email, you choose a plan or a credit pack. Starter plan: $20 a month for a monthly allowance of 20,000 OC, plus member pricing on credit packs. A launch promotion may add extra OC for a limited time; the pricing page spells out the base, the promotion and the total. Credit packs: $20, $50 and $200 one-time top-ups. Promotions, when there are any, add OC on top of the base amount for a limited time. Pack credits never expire while your account is active. What the numbers mean. At roughly $0.001 per OC, the 2...

Start your assistant setup with one boring, recurring chore

The best first thing to delegate is rarely the most impressive one. It's the boring, recurring chore with a clear finish line: a weekly summary, a morning check, a round of follow-ups. A setup order that keeps it manageable: Pick one chore you do at least weekly. Name the capability it needs: a model, a tool on your computer, an outside service. Grant the minimum access for that one job. Run it once and read the report that comes back. Then schedule it, if you want it to recur. Octuo is built for this kind of growth. Skills, sub-agents, scheduled routines and integrations can be added and changed by you, or by Octuo when you ask. Supported specialist tools on your computer, such as coding agents, can become something Octuo hands work to once you've installed, set up and authorized them, and Octuo can operate outside solutions on your behalf under the permissions you grant. Start with one. Add the second when the first has become routine. Public beta on macOS: https://www.tutuo....