Glowwellbeing

Technical notes from the 2026 curriculum

These notes are written by Glowwellbeing as an organisation. They record methods we actually run in the room. They are not a blog of opinions, and they are not case studies of named clients.

The method Rubric and artefacts

Index

Notes published in 2026

Each note is two to four short sections. Dates are publication dates. The methods were already in the taught blocks before the note was written down for this page.

Course notes

Scoring a prompt library without a beauty contest

A prompt library that is judged by whether it “sounds right” will be rewritten every time a new person joins the channel. Prompt Systems therefore scores the library as a set: each template has a brief, a version stamp, an owner, and a twenty-case test. The score on the rubric is for the set, not for the prettiest instruction.

In the lab we freeze the template, run the twenty cases, and mark factual control and reproducibility in the same sitting. If the group wants to change the instruction, they change one variable and rerun. Cases that were added because they made the template look good are taken out. Cases that look ugly — truncated input, mixed languages, missing fields — stay in, because that is the work.

The artefact we accept at Band 3 is a folder a colleague can open on Monday: templates, cases, scores, and a one-line note on what failed. A screenshot of a clever chat is not a library. We say this in the first hour because people often arrive with screenshots.

Retrieval over internal files in a one-week lab

Working Data does not ask a team to stand up a platform. It asks for a bounded folder, a decision about how files are split, and two worked failures: a question the folder can answer that the setup missed, and a question the setup answered from the wrong file. Those two failures teach more than a demo in which everything hits.

We keep the document set small enough to inspect. If you cannot name the file that supplied a sentence, the setup is not closed. Chunking choices are written in the data-readiness note so a later change is visible. Metadata that looks useful and is actually a person’s name is treated as a sensitive field, even when the file sits on an internal drive.

The note is also where we list files that must not enter the index. That list is part of the scored artefact. Teams that skip it discover the omission only when someone asks a question the folder should never have been able to answer.

Reproducibility as a taught skill

Reproducibility on our rubric is not a scientific paper. It is the ability of another person, with the files in the pack, to rerun the work and land in the same band. Chat history fails that test because it cannot be handed over and because the current instruction is usually three scrolls back.

We therefore require a log: date, variable changed, score, short note. The log looks bureaucratic on day one. By week three it is the only way to see why Tuesday’s template scored lower than Thursday’s. Facilitators mark a missing log as Band 1 on reproducibility even when the output reads well, because the next editor will be guessing.

Capstone groups sometimes want to skip the log once the prototype “works”. That is the moment we reopen the criterion. A prototype without a log has no maintenance path, and we will not call it Band 3.

Logging model calls before you have a platform

At L2 a spreadsheet is enough. Columns we insist on: timestamp, operator, input class (not the raw personal data), template version, model or tool name, output band, and a pointer to the file. If a column is missing, the sheet does not count as a log.

At L3 the same columns move next to the API call. Automation labs fail the handover criterion if secrets sit in the log, or if the log cannot be read when the author is on leave. The hand-off note to IT has to say where the log lives and who may open it. That sentence is part of Governance as well; we do not wait for a later course to invent it.

When a group already has a vendor dashboard, we still require an export the team controls. Dashboards disappear in a licence change. The evaluation set has to survive that.

Why the mechanics block is thirty-five minutes

Evening attention for abstract explanation is short. We learned this by running overtime lectures in early 2026 intakes and watching the independent lab collapse. The fix was a hard cap: 35 minutes of mechanics, including one worked failure, then the guided build. Anything that needs more time becomes a reading note issued the day before, or a second evening.

The cap also forces the syllabus to stay testable. If we cannot show the mechanism and one failure in 35 minutes, we do not yet understand how to teach it in this curriculum. GPU internals, training runs and legal doctrine all failed that test and were moved to the scope-limits list on the method page.

Facilitators are not allowed to steal the break to finish a slide. The break is 15 minutes because people need to leave the desk. A cohort that skips the break produces worse labs, which is visible on the rubric later that evening.

A data-readiness note that survives handover

The headings we require are boring, which is the point: purpose of the project, sources in scope, sources out of scope, formats, duplication, sensitive fields, redaction method, retrieval setup, known failures, owner, next review date. If a heading is blank, the note is not closed.

The “out of scope” list is the one sponsors skip. It is also the list that protects the organisation when a well-meaning colleague adds a folder of exports. Writing the list in class, on the actual files, is slower than a template with lorem ipsum. It is the work.

Capstone will not open a build week until this note exists at Band 3. That rule has delayed two in-house starts in 2026. Both groups finished with a prototype they could defend. We keep the rule.