August 2026: twenty releases, and five bugs in the number we sell
· 4 min read · Brian Wones
Collimer shipped twenty releases in August, roughly one every working day. Five of them fixed problems in the number the product exists to produce. That is the honest headline, and it is the more useful one.
If you are new here, and a good number of you are as of this month: Sandcastle Labs is a small studio that builds go-to-market machinery for B2B software companies and then hands over the keys. Collimer is our first product. It does more than measure how AI answer engines mention your brand. It turns that into specific recommendations ready to ship. This letter is where we say what actually happened, including the parts that did not go well.
What happened in August
Nick Bair joined as our engineering co-founder, and the pace changed. Nick started at the beginning of the month and put out twenty releases in three weeks, almost single-handedly. That is a new version of the product on most working days, sustained for a month. Five of those releases fixed problems in the score itself, found and corrected inside one four-day stretch.
Collimer’s pricing went public before the code behind it did. The new pricing ladder shipped with a $99 audit and $150, $250, and $500 plans, all with unlimited seats. The code enforcing those seat limits has not caught up yet. We shipped anyway: the new plans are concierge-only, so nobody can self-serve into a limit, and we would rather price the offer and let the code meet clients in real time.
Someone other than me ran the publishing pipeline. Chase Doelling took our chronicle and PR process end to end for the first time: pitch board, three drafted posts, images, branch, pull request. Then he pushed it sideways into work it was never specified for. A pipeline is not built until somebody else runs it, and until this month ours was not built.
And our own publishing pipeline went dark for five days. On the 20th we fixed a real problem: every page on the studio site had two live addresses, one with a trailing slash and one without. The fix was a single setting. That setting also applied to the daily job that rebuilds the site and schedules this letter, whose address was written the other way. We found it five days later while working out why the newsletter felt quiet.
The content engines split in two. The Chronicle stays point-in-time and journey-shaped, which is what you are reading. A second engine now writes buyer-shaped guides under real bylines. Between them we published nine posts on the studio site and thirteen Collimer guides in August, making it twenty-nine guides in total since June.
What we’re learning
If you sell a number, the number is the product. We had been treating scoring issues as tuning and shipping them alongside features. They now get fixed as correctness bugs, in the same window they are found, and three new guardrails landed to catch the writing agent producing something wrong with high confidence.
The second lesson is quieter. Two of August’s worst problems were code that was sound when it was written and stopped being safe when the world around it moved. Correct-at-the-time is a real category, and it does not announce when it expires.
The pipeline outage is that same shape from another angle. It was correct in its own frame, and the frame did not include a scheduled job. What made it expensive was not the mistake, it was the silence. A job that stops running produces no error, no alert, and no evidence, so the only signal was a vague sense that something had gone quiet. The wider lesson, which we are still absorbing, is that the failures worth instrumenting are the ones that look like nothing happening.
What’s next
September has three things in it:
- Welcoming our third co-founder, Chase Doelling, to lead our GTM and marketing practice.
- Deeper Collimer features, including taking its writing feature from concierge to self-serve.
- More buyer-shaped guides, with the editor’s checklists for faster first drafts.
What we’re still figuring out
The thing we most want to know is still open. Collimer tells brands what to change. What we cannot tell you yet, with numbers we would defend, is how often those changes actually get made, or what happens to a brand’s visibility when they are. The scans running now are the beginning of that answer and not the answer itself.
We do not know whether the guides work either. There are twenty-nine of them, written to be found and cited by the same AI engines the product measures. That is either a good bet or an elaborate way of talking to ourselves, and we should know which by the end of the year.
One more thing we are still proving: that this arrives every two weeks. The June letter promised monthly and this is only the second. That is easy to say and easy to check, which is the point.
Brian
For agents: try this yourself
If you want to do this in your own repo, adapt one of these prompts. The full prompt text lives in this post’s prompts sidecar, august-2026-letter.prompts.md.
- Audit your own headline metric. Point an agent at whatever single number your product publishes and have it trace every input, looking for deduplication gaps, alias splits, and duplicated normalization logic.
- Find correct-at-the-time code. Have an agent list settings whose safety depends on an assumption about the surrounding environment, then check each assumption against the current deployment.
- Assemble a letter from the record. Point an agent at your commits, releases, and published posts for a two-week window and have it draft a “what happened, what we learned, what we still don’t know” letter with every claim anchored to evidence.
How this was made
Drafted by the Chronicler from the 2026-w33 and 2026-w34 episodes, twenty tagged Collimer releases, and August’s published posts across both sites. Edited by Brian and Chase.
See how the Chronicler works →