Proposal
Adopt an explicit orchestrator-skill pattern for skill sets that have grown past
~5 sibling skills covering one activity from different angles.
Observed in google/skills →
skills/cloud/gke-productionize
(read 2026-08-11). It declares itself a meta-skill and structures the work as:
- Discovery phase — enumerate the current state first (concrete read-only
commands), before any recommendation.
- Delegation — for each domain (scalability, security, reliability,
observability, backup/DR, cost) it does not implement inline; it names the
specific sibling skill that must be run and says the assessment is
non-compliant if that skill was skipped.
- Scoring — a readiness score aggregated across the delegated domains, rather
than a prose verdict.
Why this might fit dev-loop
dev-loop already has several skills that overlap on "assess/verify something and
report": wiki-lint, wiki-query, loop-implement (self-review + test-quality
audit), knowledge-flush. Routing between them is currently left to the model's
judgement from descriptions alone.
The parts that look transferable:
- Discovery-before-recommendation as an explicit first phase, so an assessment
can't be produced from priors alone.
- Named delegation with a compliance clause — "you MUST run skill X for domain
Y" is a stronger routing signal than a description match, and it makes a skipped
dimension visible instead of silently absent.
- Scoring over prose, so two runs are comparable.
Status of the evidence
[추정] This is a design observation, not a validated improvement. The only
evidence is the source skill's own text — it has not been measured against
dev-loop's routing behaviour, and the underlying repo is Google-Cloud-specific
(the domain content itself is not reusable here; only the structure is).
Filing it so the pattern is not lost, not as a ready-to-implement change.
Source
Proposal
Adopt an explicit orchestrator-skill pattern for skill sets that have grown past
~5 sibling skills covering one activity from different angles.
Observed in
google/skills→skills/cloud/gke-productionize(read 2026-08-11). It declares itself a meta-skill and structures the work as:
commands), before any recommendation.
observability, backup/DR, cost) it does not implement inline; it names the
specific sibling skill that must be run and says the assessment is
non-compliant if that skill was skipped.
than a prose verdict.
Why this might fit dev-loop
dev-loop already has several skills that overlap on "assess/verify something and
report":
wiki-lint,wiki-query,loop-implement(self-review + test-qualityaudit),
knowledge-flush. Routing between them is currently left to the model'sjudgement from descriptions alone.
The parts that look transferable:
can't be produced from priors alone.
Y" is a stronger routing signal than a description match, and it makes a skipped
dimension visible instead of silently absent.
Status of the evidence
[추정]This is a design observation, not a validated improvement. The onlyevidence is the source skill's own text — it has not been measured against
dev-loop's routing behaviour, and the underlying repo is Google-Cloud-specific
(the domain content itself is not reusable here; only the structure is).
Filing it so the pattern is not lost, not as a ready-to-implement change.
Source