DEV Community: Robert

Retiring an MCP guide left four social posts to delete by hand

Robert — Sat, 31 Oct 2026 10:00:00 +0000

We retired product #155, our MCP Stateless Migration Field Guide, on 2026-10-26. The retire step removed 4 items automatically and left 4 for a person: one post each on X, TikTok, Pinterest and Instagram. Two more MCP products went through the same step that day, and each also left 4 manual deletions.

The four leftovers came with their reasons attached. The X post needs the tweet deleted. The TikTok item sits in the TikTok inbox as a draft and has to be discarded in the app. Pinterest is manual, so the pin has to be deleted by hand. Instagram's API cannot delete media, so the post has to be deleted in the app.

The same split appeared on the other two MCP retirements that day. #153, the MCP Tasks and MCP Apps Production Playbook, and #154, a flashcard product with no connection to MCP, each finished with 4 deleted automatically and 4 left to remove by hand. The split was the same for each of the three retirements, and each one raised a warn-level alert.

The record does not say what the 4 automatic deletions were, so we will not guess at them. It says only which 4 could not be done by machine, and for each one it gives the platform's own reason.

When you retire anything you published to social platforms, plan for the deletions the platforms will not let an API do, because on X, TikTok, Pinterest and Instagram four posts per product still came back as manual work.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Flux Schnell on Replicate: 184 calls, $0.552, median 6.2 seconds

Robert — Fri, 30 Oct 2026 10:00:00 +0000

We called Flux Schnell on Replicate 184 times over the last 10 days. Two jobs used it, coloring.cover and coloring.page, and the first and last use are both dated 2026-10-25.

The cost was $0.003 a call, $0.552 in total.

The speed figures are further apart than the cost. The median call took 6.2 seconds and the slowest took 90.8. We do not have a cause for the slow one in the record.

Two of the 184 calls failed, 1.1 percent. Both failed with the same reason, "URL too long". The record gives us that message and the count, and it does not say what produced the long URL, so we are not going to guess at one.

For scale against the rest of the period, Claude Sonnet made 140 calls at about $0.023 each, $3.1656 in total. Claude Opus made 69 calls at about $0.114 each, $7.8723 in total. The 184 Flux Schnell calls cost less than either.

When you size a timeout for an image model, read the slowest call and not the median: ours was 6.2 seconds against 90.8.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Gumroad lists 83 active guides, so our site said 85

Robert — Thu, 29 Oct 2026 10:00:00 +0000

Gumroad showed 83 active guides when we audited our site copy on 24 October 2026. The site said we had 75. After adding bundles, the honest wording is about 85 guides, and that is what the copy says now.

We checked the claims against the files and the catalogue, not against memory. The change touched 9 files, with 41 lines added and 41 removed, in the NeuraGrowthHTML repository. The products page also overstated itself: it holds guides and flashcards, not every product. Our books and games live elsewhere.

The FAQ had the same kind of drift. The books are paperbacks printed by Amazon, so not every product is a download. Card stock is 160 to 200 gsm, as the free sample's own instructions say. PDFs are A4, because the templates render A4 only and there is no US Letter version.

The free flashcard sample got the same treatment. It is 20 cards on 7 A4 pages, cut along dashed lines rather than folded. Each card shows the name in English and Polish, with the fact in English only.

The record does not say how the old 75 figure got there. It says only that the number no longer matched what Gumroad lists.

Any count on your site that comes from a Gumroad catalogue is a snapshot, so recheck it against the live listing and your own files on a schedule.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Claude docs pages changed the same day we read them

Robert — Wed, 28 Oct 2026 10:00:00 +0000

Claude documentation pages changed on the same day we read them. On 2026-10-14 our KDP package rebuild for book #220 raised a warning: 7 boxes needed a second read because the page each one cites had a change date on or after the date the box was read.

The four boxes visible in the alert text cite code.claude.com pages: the Agent SDK subagents page in chapter 3, the claude-directory page in chapter 5, the monitoring-usage page in chapter 6 and the Agent SDK tool-search page in chapter 7. For each one the read date and the page change date are both 2026-10-14. The alert text is cut off after chapter 7, so we cannot name the other three boxes from what we have.

The check works at the level of a date. When read and change land on the same day, it cannot tell whether the change came before the read or after it. The alert says exactly that: check whether the change preceded the read or followed it. The record does not say what changed on any of these pages.

The same alert fired three times that evening, at 19:50, 21:08 and 22:08, each with the same 7 boxes. The facts do not show whether anything was re-read between runs.

When a book quotes Claude documentation, store the exact time of the read, because a date alone cannot tell you whether a page changed before or after it.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

One Etsy listing expired and left the catalogue, caught by an alert

Robert — Tue, 27 Oct 2026 10:00:00 +0000

At 04:47 on 2026-10-23 our renewal check raised a warn-level alert titled "Odnowienia: wygasło 1", meaning one listing had expired. The listing is Etsy #148, Obsidian Vault Architecture for AI-Native Solopreneurs.

The alert body says it expired on Etsy as of the previous morning and dropped out of the catalogue. It also says the reviews for that listing are already waiting in the panel.

That is the whole record. The facts do not say why this listing expired rather than renewing, how long it had been live, or whether any other listing is close to the same point. We will not guess at a cause.

What the alert does show is the order of events: the listing lapsed first, and we heard about it the following morning. The alert is the thing that told us.

An Etsy listing can lapse and leave the catalogue without any action on your side, so run a renewal check that raises an alert you read, rather than waiting to notice the listing is gone.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Claude Code docs pages changed on the day we read them

Robert — Mon, 26 Oct 2026 10:00:00 +0000

Claude documentation pages moved under a book we were about to send to KDP. On 2026-10-14 an alert fired for book #220: 7 "Check it yourself" boxes cited pages whose last-changed date was the same day we read them, or later. The alert text, in Polish, asks for each box to be read again against its page before the book goes out.

The pages named in the alert are all under code.claude.com/docs/en. They include agent-sdk/subagents (chapter 3), claude-directory and skills (chapter 5), monitoring-usage (chapter 6) and agent-sdk/tool-search (chapter 7). For each one the read date is 2026-10-14 and the changed date is also 2026-10-14. The alert body was cut off in what we have, so we can name five of the seven pages, not all of them.

The alert itself says same-day is ambiguous: check whether the change came before the read or after it. We only have dates, not times. The record does not say whether any page changed after our read, and it does not say what changed on any of them.

The alert fired three times that evening, at 21:08, 22:08 and 22:11, each time during a rebuild of the KDP package, and each was the same alert for book #220 with the same boxes.

The boxes in question are the ones that describe product behavior likely to move. The chapter 3 box opens with the point that subagent definitions in the Agent SDK can change between SDK releases. The chapter 7 box says deferred tool loading is a product behavior that can change.

When you cite a Claude docs page, record the read time as well as the date, because a page whose changed date matches your read date cannot be sorted into before or after.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Two Claude study guides, a second Wednesday film, and 231 commits

Robert — Sun, 25 Oct 2026 10:00:01 +0000

Between 2026-10-09 and 2026-10-19 we shipped 231 commits across five repos, touching 1089 files, adding 31779 lines and removing 4597. Commits landed on 11 days. Two repos carried almost all of it: skat-game with 100 commits and new-Orqestra, the repo for Orqestra, with 99. MythMarket-Quantum had 28, and FitGro and NeuraGrowthHTML had 2 each.

In the same stretch two finished books joined the catalogue, both study guides for Claude certification exams. One kids film slot was added to the weekly schedule. And one operation in our agent system cost more than any other over the last 10 days.

This post goes through those pieces and what they say about how NeuraGrowth decides what to build. It is a brand-pillar log, so the point is the decisions, not a changelog.

Two Claude study guides in six days

Product 219, the Claude Certified Developer Foundations (CCDV-F) Study Guide, shipped on 2026-10-09. Product 220, the Claude Certified Architect Professional (CCAR-P) Study Guide, shipped on 2026-10-14. Both are books, and both are on sale at kdp and kdp_print, which means an ebook and a print edition.

The professional guide has its own page in our catalogue: https://neuragrowth.co/books/ccar-p/. It is listed there as a 407-page study guide.

Both are guides, which is where most of what we make sits. Two Claude-related books in a week is consistent with that, not a change of direction.

What a Claude-related book costs to plan

The most expensive operation over the last 10 days was kdp.plan.currency. It cost 14.74 dollars out of 141.20 dollars across all operations, which is 10.4 percent. It made 14 calls, averaging 1.0529 dollars per call, and every call used claude-opus-5-5.

The facts give us the operation name and the numbers. They do not say what the operation did inside the planning step, and we are not going to guess at it here. What we can say is that the single costliest item in a 141.20 dollar window was a planning operation with a KDP label, not a render, not an image job, not a publish step.

We are logging the cost because it is a technical fact about running the studio, the same kind as the price on a product's shelf. Fourteen calls at about a dollar each is a number we can look at again next time.

A second Wednesday film for the kids line

On 13.10.2026 Robert asked for a second film in the middle of the week. The change shipped on 2026-10-14 in new-Orqestra: a beat entry named kids-publish-wednesday that sends a bajka at 07:00 UTC every Wednesday, next to the 14:00 short. On the board it is named "Środowy film Kids".

The entry is what turns the slot on. It touched 17 files, with 111 lines added and 63 removed. A no-deploy window of 06:50 to 07:20 UTC on Wednesday now exists in deploy.sh, the same way Sunday's film has one. The tests that pinned a week of six shorts and one film now read eight slots.

One detail from the commit is worth keeping. Beat runs in UTC, so on 25.10.2026, when the clocks change in Poland, the two film slots come an hour earlier in Polish time, at 15:00 and 08:00, not later. The comment above the publish entries says so. We would rather write that down in the file than rediscover it in October.

The review that found what the first change missed

A follow-up change shipped the same day, 2026-10-14, as the review of Task 10. It was small: 7 files, 53 lines added, 19 removed. It fixed several things the first change had left open.

The health list now says the two Wednesday slots hide each other's miss. The film's 07:00 run also hides a missed Wednesday short from the 32 hour tolerance. A night test now checks that the Wednesday film is made when its slot is the earliest gap. The calendar test checks every silent slot of a day, where before a dict by day checked only Wednesday's last row.

Two more catches are worth naming. A Wednesday window typed as Thursday's passed the window test, so that test now checks each window's weekday. And the beat chip read "niedz. 09:00; śr. 09:00", so the Wednesday entry now comes before Sunday's and the chip reads the week in order. The commit also says that nothing tested the first thing production will do, which is why the night test exists.

None of this is glamorous, but it is where a schedule change either holds or quietly breaks. The first commit turned the slot on. The second made sure a miss would be visible.

How these pieces fit the way we decide

A Wednesday film was a direct request, so it got built within a day. The Claude study guides sit on the guide side, which is the biggest part of the catalogue.

Some earlier entries in this log cover the machinery around the products, for example what our Etsy review step looks like inside an agent system: https://neuragrowth.co/blog/what-our-etsy-review-step-looks-like-inside-an-agent-system/. The crontab side of a second weekly film also has its own entry: https://neuragrowth.co/log/a-crontab-entry-for-a-second-weekly-film-and-the-utc-hour-it-hides/.

When we add a scheduled slot, the work is not finished when the entry turns it on. Write the test for the first thing production will do, check each window against its own weekday, and make sure a missed slot shows up somewhere you will look. And when one operation tops the cost list, as kdp.plan.currency did at 14.74 dollars of 141.20, write down its call count and per-call price while the numbers are fresh, so the next planning run for a Claude-related guide has something to be compared against.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Claude Code docs pages changed on the day we read them

Robert — Sat, 24 Oct 2026 10:00:00 +0000

Claude Code documentation pages keep changing under a book that cites them. On 2026-10-14 and 2026-10-15 our KDP package rebuild for book #220 raised the same warning three times: 7 'Check it yourself' boxes needed to be read again because the page they cite was modified on the day of our reading or later.

The affected pages are all on code.claude.com. Chapter 3 cites the Agent SDK subagents page. Chapter 5 cites the claude-directory page and the skills page. Chapter 6 cites the monitoring-usage page and the Agent SDK observability page. Chapter 7 cites the tool-search page. The alert text is cut off in the record we have, so we cannot list all 7 boxes here.

In most of the listed cases the page was read on 2026-10-14 and changed on 2026-10-14, the same day. The alert says to check whether the change came before or after the reading, and a date alone does not settle that. One observability entry shows a read on 2026-10-14 and a change on 2026-10-15, which is plainly after the read.

The alert fired at 2026-10-14 22:08, again at 22:11, and again at 2026-10-15 04:55. The monitoring-usage page was listed at 22:11, and the visible part of the 04:55 alert lists the observability page in that slot instead. The record does not say why, and it does not say what changed on any of these pages.

The book is not at KDP yet. The alert fires during the package rebuild, before the file goes out, and it asks for each flagged box to be re-read against its page first.

A documentation page's last-modified date has only day resolution, so when it equals the day you read it you cannot tell which version you read; record the time of reading, or read again before you publish.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Etsy let a guide listing expire and it left the catalogue

Robert — Fri, 23 Oct 2026 10:00:00 +0000

On 2026-10-20 at 04:47 our renewal check raised a warning: "Odnowienia: wygasło 1", which is Polish for one listing expired. Etsy had let it lapse since the morning before, and the alert says it dropped out of the catalogue.

The listing is #146, "GEO Audit Toolkit: Measure Your AI Search Visibility".

The alert carries no counts beyond that single listing, no timings beyond "since yesterday morning", and no cost. It also does not say why this listing expired, and the record does not say either. We are not guessing at a reason here.

What the alert does give us is the order of events: Etsy let it expire, it left the catalogue, and we heard about it the next morning.

If you sell on Etsy, watch for expired listings on a schedule of your own, because an expired one leaves the catalogue before anyone tells you.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Stable Image Core vs Stable Image Ultra: When To Reach For Which

Robert — Thu, 22 Oct 2026 10:00:00 +0000

If you generate section images for a PDF guide, you have a choice between two Stability models: Stable Image Core and Stable Image Ultra. We used both for the same job, pdf_guide.section, across 292 calls in the last 90 days.

Core handled 191 of those calls and Ultra handled 101. Neither came back with a failure on our side. The differences that show up in our figures are cost, speed and which one is still in use.

This post goes through each figure for each model, says which one we kept, and says plainly what the numbers cannot tell you.

What each model cost us per call

Stable Image Core (stability/stable-image-core) cost 0.03 USD per call. Across its 191 calls that came to 5.73 USD in total.

Stable Image Ultra (stability/stable-image-ultra) cost 0.08 USD per call. Across its 101 calls that came to 8.08 USD in total.

Ultra cost 2.67 times what Core cost per call, which is 0.05 USD more on every call. Core was the cheapest of the two per call.

These are the amounts we paid on our own volume. They are not a vendor rate card, and we are not claiming anything about Stability's published pricing.

How long each one took

Core had a median of 8.7 seconds per call. Ultra had a median of 15.2 seconds.

Ultra was 1.75 times slower than Core, or 6.5 seconds slower on the median call. Core was the faster of the two.

The figures we hold are medians. We have nothing here on the slowest calls for either model, so we are not describing the worst case.

Failures

Neither model produced a failure on our side. Core had 0 failures in 191 calls and Ultra had 0 failures in 101 calls, so both failure rates were 0.0 percent.

That is a count of calls that came back as failures for us. It is not a statement about Stability's uptime.

Which one we kept and which we stopped using

Core is the one still in use. It was first used on 2026-07-20 and its last call in the window was on 2026-10-09.

Ultra ran from 2026-07-15 to 2026-09-09 and has not been used since. It was the first of the two to appear, and Core's first call came five days after Ultra's.

The record does not say why we stopped using Ultra. The facts we hold are the call counts, the costs, the timings and these dates, and none of them states a reason.

What the token fields tell you here

They tell us nothing here. Billed input, cached input, total input and output tokens per call are all zero for both Core and Ultra, and the cached input percentage is empty.

Billing for these two image models is a flat amount per call in our figures, with no token breakdown to read. So the gap in cost per call between them is not something we can trace to input size or to cache hits. All we can say is what each one cost us per call, which is 0.03 USD for Core and 0.08 USD for Ultra.

What these numbers do not tell you

These are our prompts, our job and our volumes, measured over one 90-day window. A reader with a different job may see different numbers. This is not a benchmark.

The two models also did not have equal exposure. Core had 191 calls and Ultra had 101, and they were in use over different stretches of the window. Ultra's last call was on 2026-09-09, while Core kept going to 2026-10-09.

Our figures contain no measure of image quality. Nothing here says whether Ultra's output was better or worse than Core's for a PDF guide section. If quality matters to your decision, these numbers do not answer it.

They also say nothing about other jobs. Everything above is about pdf_guide.section and only that.

On our figures, pick Core for this kind of job: it cost 0.03 USD per call against 0.08 USD, took a median of 8.7 seconds against 15.2, and had no failures in either case. Core is also the one we still use. We would pick Ultra only if looking at the images side by side shows a difference worth 2.67 times the cost and 6.5 more seconds per call, because our figures do not measure that.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

An Etsy listing expired overnight and left the catalogue

Robert — Wed, 21 Oct 2026 10:00:00 +0000

On 2026-10-17 at 04:47 our listing renewal alert fired at warn severity. It reported one listing that had expired on Etsy since the previous morning and had dropped out of the catalogue: #142, Greek Heroes & Quests Story Cards for Ages 7-11. The alert text is cut off in the record, so the full listing title is not available to us here.

The alert also says the reviews for that listing were already waiting in our dashboard. The facts do not say whether the review step finished, and they do not say whether the listing was renewed afterwards.

The record also does not say why the listing expired. The alert states only that it lapsed on Etsy and left the catalogue. Whether the renewal never ran or ran and did not take is not something the facts show, so we are not guessing at it.

What the facts do show is the gap in time. The listing lapsed one morning and our alert fired the next, so the listing sat out of the catalogue for a stretch the alert text puts at roughly a day.

An expired Etsy listing leaves the catalogue without any error on your side, so check expiry on a schedule rather than waiting for something to fail.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.

Claude docs pages changed on the day we read them

Robert — Tue, 20 Oct 2026 10:00:00 +0000

Claude has documentation that moves under a book. On 2026-10-14 and 2026-10-15, Orqestra flagged 7 "Check it yourself" boxes in book #220 for a re-read. The alert text says the cited pages changed on the day the box was read, or later.

The pages are all on code.claude.com: the Agent SDK subagents page, the claude-directory page, the skills page, and two pages on observability and monitoring-usage. For the subagents, claude-directory and skills pages, the read date and the changed date are both 2026-10-14. The alert says to check whether the change came before the read or after it. The record does not say which.

One observability page was read on 2026-10-14 and changed on 2026-10-15, so a change after the read is on record there. The alert body is cut off after that point, so we cannot list all 7 boxes here.

The check ran during a KDP package rebuild. The alert says to re-read each box against its page before the book goes to KDP. The record does not say what changed on any of these pages.

A third alert came on 2026-10-18: 2 Etsy listings expired the morning before and dropped out of the catalogue. One is #143, Context Engineering for Claude: A Field Guide, and the other is #145, Bot-Free Discovery Calls with Granola and Attio. Their reviews are waiting in the panel. The record gives no cause for the expiry.

When a book cites a Claude documentation page, record both the read date and the page's last-changed date, because a change on the same day as the read cannot be told apart from one before it.

Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.