Open questions and honest tensions
Seven build-fresh items in cost order, six open questions, five tensions. The brief pack is explicit that publishing these unresolved is the point, not a gap in the site.
Open questions
What is a skill version?
Semver assumes behavioural compatibility claims; a skill's behaviour depends on the model reading it. Does 1.2.0 mean anything when the interpreter changes underneath?
Unaddressed anywhere in the corpus. The intent-over-capability claim on the thesis page makes it harder, not easier.How do you test intent?
The thesis demands CI for skills; no one has said what a failing test looks like for “an opinionated way of approaching something.”
Evals are the obvious answer, and the estate has none.Who arbitrates triggering conflicts?
The governance rules on the authoring page say skills compete for triggering — but with no registry, nothing detects that two skills claim one intent.
The catalogue could: a build-time overlap check on trigger vocabularies is feasible today and has not been run.Is the base vault the skill, or its source?
If the SKILL.md is a projection of the vault (per the graph page), which one is versioned, reviewed and sold?
The economy briefs assume the vault; the shipped practice is SKILL.md-only.Does the permission set belong to the skill or the invocation?
The 18 June brief wants moment-of-authorisation — which implies the invocation computes it — while the frontmatter proposal declares it statically.
Both, probably: a declared ceiling, a computed grant within it. Needs a design.Would Tessl-style aggregators accept vault-backed provenance?
The monetisation model in the economy page rides platforms whose scoring is exactly the “gameable popularity” the vault position rejects.
A real conversation exists; the answer is not recorded anywhere in the corpus.Honest tensions
- The theory criticises the practice, and both are ours. Every shipped skill is the “static photograph” the projection paradigm dismisses. Publishing that is the credible move; resolving it is a research programme.
- One green cell. The lifecycle table on the thesis page scores the estate’s skills against the estate’s own checklist and passes only documentation. This site leads with its best artefacts and their worst scorecard.
- Skills earn their keep on distinct tasks — and this estate ships eight, while the theory imagines thousands. The governance rule that keeps quality up is in tension with the economy that needs volume.
- The intent claim cuts both ways. If skills describe intent in English, then model drift is dependency drift — and there is no lockfile for the interpreter. Q1 is not pedantry; it is the hard version of the thesis.
- The strongest week of thinking produced the least code. 1–4 June: roughly 35,000 words, zero commits to the skills library in that window. This site is partly a mechanism for closing exactly that gap.
Build-fresh items, in cost order
| # | Item | Why |
|---|---|---|
| G1 | Resolve the use-sg-playwright duplicate | Two copies, one carrying a claim the code contradicts. One repo owns it; the other imports. |
| G2 | A permissions field in SKILL.md frontmatter | Declared, not enforced. Starts the identity design with a day’s work. |
| G3 | One eval for one skill | Zero exist. The scoring differentiator has two of its three legs; this is the third. |
| G4 | A version field per skill | The cheapest possible start on the package thesis. |
| G5 | The authoring guide | Done — this site publishes it. |
| G6 | A skill review process | The Feb 2026 trio proved it can be done; nothing since. Even a checklist referencing that precedent is a start. |
| G7 | The registry v0 | Done — the catalogue, generated, is exactly that. |