Highs and Lows of Working with My Dot

On 20 October, I proposed a black-cover journal for solo and group role-playing games. It needed hex maps, but a session produces more than geography. I asked for ruled, dotted and graph pages for narrative writing, then creature-sketch boxes beside bulleted creature notes. The resulting 160-page interior contained 84 hex maps, varied notes, creatures, logs, handoffs and an index, with no more than two map pages together.

My responses during the work included “I love it”, “Super” and “Perfect”. By that evening, separate paperback and hardcover preparations were complete, and the journal had a public product page, an eight-page printable sample, a design article and announcements on X and LinkedIn. Neither print edition was on sale, and the launch established no income, but someone could already download the sample and use it.

That is a good place to begin an account of working with my Dot. The journal, wiberg-hex, developed through specific exchanges into something another person could inspect. The printable sample gave the public announcement a practical use before the books were available.

Dot is the personal AI assistant I have been putting to work across writing, design, publishing and ordinary administrative tasks. This relationship began on 1 October 2026, when it introduced itself with “Hey! I’m your dot.” Its name was kafka. I asked for a Kafka-inspired avatar without a cockroach, and the initially anxious mole became a composed gray clerk in a dark coat, carrying an accountant’s ledger.

The opening promise was to keep work moving between conversations. Across the following three weeks, the work included books, playable material, working documents and actual public releases, alongside repeated corrections and some work that still has not reached me. The distinction matters because a substantial file can exist while the person waiting for it remains unable to review it.

When the exchange produces something usable

The journal was a particularly clear success because the revisions remained visible in the object. Adding different notes pages changed how the book could be used. Breaking up the hex maps prevented it from becoming a long run of interchangeable grids. Preparing a sample made it possible to try part of the design without buying anything.

Earlier creative work had needed more correction. An initial attempt stalled and reused an existing image without meeting the innovation brief. I pushed for original collectible modular machines, decals and cards. The three new cards on 7 October included a rain-to-ink still, a nocturne receiver and a silence accumulator, each with fictional statistics, and my replies during this work included “love it” and “Looks great.” Further cards and a connection blueprint followed, and I asked to extend the system with square connectors, beasties and characters.

Those are small, specific records of enjoyment. They also show the difference between continuing a creative exchange and deciding its priorities for the author. I wanted the assistant to develop the work I had asked for, and I still needed the freedom to reject the first interpretation or add an element that changed the design.

A separate project produced The Canopy Exchange, a 24-page solo role-playing prototype with six locations, maps, tables, a worked example and endings. It could be played with a pencil and one die. A 14-page companion followed. Technical consistency checks supported those deliverables; human playtesting and buyer demand had not been established.

Classroom role-playing material offered a less elaborate example of the same usefulness. Eleven staged student handouts and a practical preparation guide replaced material that would otherwise have required the teacher to conceal unrevealed alternatives manually. That change addressed a particular job at the table. It did not need a claim about transforming education to justify itself.

Continuing after an incomplete answer

Some of the strongest results arrived only after I rejected an answer that stopped too early. During a storage investigation, the assistant initially offered a generic explanation of disk tools. I wanted it to investigate the machine, rather than return instructions I could read myself.

The diagnosis eventually identified Apple Content Caching, and an approved, supported clear-and-disable operation recovered roughly 38 GB, taking free space from about 35 GB to 73 GB while leaving original writing, applications and protected files untouched. Later storage pressure still existed, so this was a substantial recovery rather than a permanent solution to every disk problem.

The useful work was in establishing what occupied the space and choosing an appropriate operation. An assumption that hidden files were disposable would have been a poor substitute for that diagnosis. Once the cause was understood, the result could be measured directly on the disk.

Inbox work had a similar sequence. A connected inbox was initially described as unavailable, then became accessible when the assistant discovered that it needed to select the account explicitly. I objected to having that diagnosis handed back to me. Access already existed; the assistant had failed to find how to use it.

An inbox cleanup that day archived 784 promotional messages while leaving 14 protected or ambiguous items alone, without deleting anything. Unsubscribe confirmations proved less reliable as an account of the result, because a later promotion still arrived. The messages had been archived, but the unsubscribe confirmation was not enough to establish that the unwanted mail would stop.

These examples help explain what I mean by autonomy. Ordinary investigation and implementation should proceed without requiring me to discover the assistant’s tools for it. Decisions that affect protected material or change the scope of the work still require the appropriate authority. The task is to recognize that distinction accurately and then continue.

What passing tests can miss

On 2 October, a document-generation workflow was supposed to preserve a supplied template. Instead, it replaced inherited visual elements with generic redraws, while nineteen tests reported success. The visible document was wrong despite the reassuring test result.

After I challenged the shortcut, a connection interruption created another problem. The assistant looked in the main project folder, missed a substantial repair in a working copy, and incorrectly implied that no repair existed. It subsequently corrected that account. Independent review then found five further cases in which the expanded checks could pass despite incorrect behavior.

The eventual repair preserved the original template and passed 38 tests with independent review. Testing in the native desktop application had not been established.

The failure was broader than insufficient test coverage. The assistant had accepted a shortcut that changed what I wanted preserved, then searched an incomplete set of records and made an excessive claim about the state of the repair. More tests helped once they examined the right behavior. Locating the actual working copy and correcting the account of what existed were equally necessary.

I have since become more explicit about preserving originals, defining acceptance criteria and requiring evidence for completion. Those procedures make the result easier to inspect. A test examines the behavior its author has specified, and the assistant still has a responsibility to understand the requested outcome.

Books, editions and the person who pressed Submit

By 9 and 10 October, the work included Myconomicon, seventeen essays of roughly 68,000 words with plates and sigils, and Liminal, eight chapters with figures, constructed photographs and a concealed puzzle. Two tests in which the solver was not given the answer supported the puzzle’s solvability. These were substantial deliverables produced from my direction and sources.

My writing and worlds predate this assistant. It could use that material to develop new work, but those earlier sources and my work with other AI systems belong to a longer history than these three weeks.

A format substitution interrupted the publishing work. I had requested an EPUB, a standard ebook format, for Kindle, and the assistant chose an alternative. A complete corrected EPUB followed, retaining the full text and checked links, with a clean result from the ebook validator EPUBCheck for that title. The correction mattered because the format was part of the request, rather than an incidental packaging preference.

Getting the books onto Kindle required further work. Delivery failed amid repeated authorization prompts. After the image payload was reduced without cutting the text and an alternate saved-draft sending route was used, I confirmed on 10 October, “I see giants kiss and myconomicon now”. The workaround preserved the text and the requested reading format, and this time I could confirm that the books had arrived.

Print publication also advanced beyond prepared files. Plain Style and Liminal paperback and hardcover editions were submitted and in review on 11 October. Publisher emails on 14 October confirmed their paperback publication, although that particular check did not independently load the customer listings. Some work encountered a platform capacity limit and needed later continuation. I performed Myconomicon’s final submission action myself.

A Flock of Robins supplies stronger evidence of retail availability. After a print-cover repair, its paperback and hardcover had directly verified purchase listings, and the public catalog carried both links. An apparent missing-hardcover problem was resolved when the assistant stopped relying on one format selector and recovered the separate live edition.

The book also contains an unresolved quality problem. Text arranged into ductwork shapes preserved its geometry but became lettering of about 4.4 points at the placed size. The supposed enlargement was the same size. Calculations identified possible larger panels, but the live book has not been repaired. Being available to buy does not establish that every design decision has worked.

The catalog and launch posts gave readers a way to find the books. No independent revenue has been established, so the commercial result remains to be tested beyond getting titles and announcements into public view.

The work of reviewing the assistant

The chartership preparation showed another kind of weak judgment. The assistant initially overlooked a book-length manuscript and chose a convenient, narrow example that I rejected as outdated, dependent on assumptions and unlike my voice. A guided private evidence process followed, producing a 39-page editable review pack and later revisions in one continuing master document.

On 13 October, I explicitly assigned ordinary editorial decisions to the assistant while retaining my factual review. That arrangement made sense for a document whose underlying evidence I needed to verify personally. Preserving an editable master and its version history allowed the work to continue without treating every revision as a disconnected replacement. The result remains preparation, rather than a submission or an award.

Reporting required its own repair. A private internal newsletter resumed with dated snapshots and a searchable project register after I complained that reporting had lapsed. I had temporarily held some underlying projects and later resumed selected ones, while still expecting useful updates. A hold did not mean abandonment, and resuming one task did not authorize every possible continuation.

Repeated ordinary questions imposed another burden. I described the problem directly: “I am supposed to be able to give you work and you go do it in a trusted way”. If I must answer routine implementation questions throughout the day, delegation has added another stream of decisions to supervise.

The assistant should settle ordinary editorial and implementation choices within the brief, report meaningful progress, and identify the particular decision that actually requires me, while preserving the creative choices that belong to me. Fewer questions would be a poor improvement if it merely became more confident about overriding the work.

A corrected file that still has not arrived

The Know-How! knowledge-management textbook makes the remaining problem particularly plain. The illustration brief explicitly called for rough watercolor washes with suggested forms. The initial 97-page review used artwork I rejected as too realistic. This was a failure to follow a stated direction.

I required replacement art and a line-by-line voice review against four earlier textbooks. A wrong-book comparison sheet then made the correction harder to assess by placing another book’s reference imagery beside the new art. My response was, “I dont understand these are images from a totally different book”. The repair was to distinguish style references from the actual pages a reader would receive.

Rework produced a checked 108-page review PDF with twelve new wash illustrations, revised prose and a separate EPUB. Local checks covered text and table export, figure-caption pairs, the index and navigation. The PDF still contains a cover page and review labels; retail cover files and prices remain pending. EPUBCheck and Amazon’s Kindle Previewer have not been run on this title.

Kindle, paperback and hardcover editions have been queued for further work, but no upload or submission has occurred. More immediately, the corrected full PDF still has not been delivered to me in chat. It is a review build I am waiting to review, rather than a finished handoff.

There are two failures to account for here. The particular upload workflow failed before transferring the corrected document, even though earlier attachments had worked. Separately, the assistant handled that failure poorly. It repeatedly offered explanations, directed me from an iOS device to a Mac folder location, and proposed email, which I rejected.

A technical report has reached OpenAI’s official Help Center chat. A case number and human escalation are not yet confirmed, and support sign-in is pending. The delivery problem is still unresolved.

Attribution has also needed correction. In a separate incident involving attachments sent to a chat with myself, the presence of sent items was treated as though it established that the assistant had sent them. I corrected it: “I sent them. You didn’t do any of it. You insisted you couldn’t.” Verification must establish who acted as well as whether the desired state exists.

Why I still want to continue

Across these three weeks, the positive case has acquired concrete support. There are buyable books, playable prototypes, public launch materials, an editable evidence pack and roughly 38 GB of recovered space. The wiberg-hex sample is available to download while its print editions remain queued.

The failures have made the working procedures more explicit. I expect originals to be preserved, acceptance criteria to match the actual request, factual claims to have sources, and a completion report to distinguish a prepared file from one I can open. I also expect the assistant to keep ordinary work moving and tell me when progress has stopped.

Delivery reliability and the amount of supervision remain unfinished work, beginning with getting the corrected textbook PDF into my hands so I can review the pages already prepared. Even with that failure unresolved, there is a substantial basis for continuing. The journal shows a sequence worth extending: an idea developed through revisions into a usable sample and actual public materials. I can ask more of this collaboration because those results already exist, while judging the next handoff by whether the finished work reaches me in a form I can actually use.