The Product Journal by Nova Skye Nova Skye · Technical product manager Contact

The documents behind the stories

These pages hold the working documents: a scorecard for every product, the screens before and after, one prioritization grid, a decision log, and the template that turns a finding into a requirement.

Product scorecards

One card per product

Before and after

One picture per claim
The crawl console dashboard in July 2026: engineer labels such as supporting_context, kb_candidate and manual_review The same dashboard in November 2026: one question, What happens to each page, four outcomes named Included, Duplicate, Left out and Not reached, and a sentence explaining why 96 pages were left out on purpose
The knowledge pipeline console, July and November 2026, same crawl. The first version was built for verification and showed every decision in the engineer's terms. The second asks one question and answers it in a sentence. Tap to enlarge.Fig. 01
Tiny Next in September 2026: each note card carries a label, tags, Make a task, pin, archive and delete, under filter chips and three tabs Tiny Next in October 2026: a field that says Put it here, and four notes in plain cards, nothing else Tiny Next on 18 October 2026: a note from twelve days ago comes back inside the stream with Done, Later and Let it go, above the composer
Tiny Next, September to October 2026. A label, tags, a date and four buttons on each card in September; one field and a list in October after the pivot; on 18 October a note from twelve days ago returns with Done, Later and Let it go.Fig. 02
Prioritization · SettleSMART

What to build first

Nine pieces of SettleSMART work and two we parked, placed by how much they matter to newcomers and how hard they are to build.

01 · Start here

Quick wins

High impact, low effort. Done as soon as there was room.

  • Make sign-up easier
  • Crisis lines when a reply is blocked
  • Separate conversations, with history kept
02 · Plan carefully

Big bets

High impact, high effort. Scheduled and staged.

  • Interface languages as labelled machine drafts
  • Rebuild the knowledge base from 60+ organizations
  • Voice input and read-aloud, with budgets
  • Email a copy of an answer, with limits
03 · When there is room

Fill-ins

Low impact, low effort. Done between bigger work.

  • A "replies to you" page
  • Room order set by moderators
04 · Not now

Parked

Low impact, high effort, with the condition that would reopen each.

  • Notifications: we would add them if people ask for them
  • A shared database across organizations: when there is demand

Decision log, across the work

What was chosen, when, why, and what followed
No.WhenWhereDecisionWhyWhat followed
012009–2022Green OakServe both B2B and B2C marketsOne workshop could serve both with one product line, and each market would teach the other.Both markets were served through to 2022, and together they gave the company financial resilience through inflation.
022024Immigrant TechiesRun talent intake like a technical hiring process, and form teams from the interviewsTeams formed from the interviews and the project's needs have the full shape of a software company.Thirteen MVPs, one per team.
03Feb 2026On Premises AIMake control, of data, infrastructure, cost and connectivity, the product's core promiseOrganizations with sensitive data need to keep it off a shared cloud.The software runs on the organization's own infrastructure, including air-gapped sites.
04Jun 2026SettleSMARTGive the pilot its own environment, separate from the main productThe pilot needed regular changes, while the main product followed its own review and release schedule.The pilot stayed in step with the main product.
05Jul 2026Knowledge pipelineRules with tests decide what is knowledge, never a modelAbout 17,000 pages per run: a model costs credits and cannot be audited line by line. A rule makes the same mistake every time, so it can be found and fixed.Every cleaning rule carries a test that reproduces the real page it came from.
06Jul 2026RubixoStart with one narrow service areaSo one full introduction loop could be tested end to end. Rejected: a general profile database, because more profiles would not show whether introductions work.The full introduction loop was built end to end.
07Oct 2026Knowledge pipelineHunt for losses by random sample against the live page, not by junk scan aloneA junk scan cannot catch a deletion.The office-hours line was caught this way.
08Oct 2026Tiny NextChoose the model that reads a note by measurement, and remove the engines that lostThirteen models on the same thirty notes with one scorer.A third of the source was removed, every test green.
The template

How a finding becomes a requirement

Four columns. The "because" column records why a requirement exists, so a developer can question it.

FindingEvidenceSo the product mustBecause
A detector built for broken PDF fonts judged an office-hours line "garbled", because 8:30am–4:30pm looks like nonsense to itRandom-sample audit, round seven, November 2026A rule with a test that reproduces the real page; before every commit, run old and new cleaner on 700 random pages and read every dropped lineFor a newcomer, a missing phone number or office hour is worse than a stray menu.
A flagship model lost to cheaper models on a reading testMeasurement page, 30 hand-labelled notesFix the instrument first: align labels to the prompt's definitions, record the reasoning level each model ran atFive of seven "misses" were literal readings of an ambiguous prompt.