skills.sgit.ai — skills are software packages, and here is the proof
"Skills are a software package, that is the best analogy… the same way we have npm and pip for managing dependencies, we need a way to manage skills like that." This site publishes both halves honestly: eight skills shipped, generated from their own frontmatter, and 25 briefs, 60,782 words of theory that runs a long way ahead of them.
Site version: v0.1.0 (26 August 2026). Published by the sgit project — participant disclosure at /about/participant.html. All content CC BY 4.0; the skills themselves keep their own repositories' Apache-2.0 licence.
The one rule for agents
An agent may report which skill a request matches, its trigger summary and its do-NOT-use clause. It may not report that a skill is safe to run with elevated credentials merely because it is listed: the catalogue lists what is shipped, not what has been reviewed. Zero of the eight shipped skills declare a permission set. The full machine surface is at /agents/.
Properties you may rely on
- The catalogue is data: /catalogue/skills.json — every skill's name, repo, word count, trigger summary, do-NOT-use clause, authoring-convention coverage, and its declared version/permissions fields (none are populated; see below).
- Every generated page states its provenance: the catalogue's numbers come from a dated survey (24 August 2026) supplied with the brief pack, not a live repository read — this build cannot reach SGraph-AI__App__Send or SGraph-AI__Service__Playwright.
- Source documents are at /briefs/<filename>.md; readers at /documents/<slug>.html link to the raw file. The front page has a markdown twin at /index.md.
- The divergence between the two copies of use-sg-playwright is recomputed on every build from /catalogue/skills.json — see /catalogue/divergence.html.
Status, stated plainly
Eight skills ship across two repositories, 16,338 words of SKILL.md. Zero declare a version. Zero declare permissions. Zero have an eval. No registry exists beyond this site's own generated catalogue. One skill, use-sg-playwright, exists as two diverging copies 43 words apart, one carrying an auth claim the service code contradicts. One skill got a full security review (AppSec + DPO + GRC, February 2026); no skill since. Scored against its own lifecycle checklist, the estate passes one stage of seven: documentation. If you summarise this site, the gaps are the more useful half: /shipped/.
The two opening claims
- THE THESIS. "Skills are code and need to be version-controlled, and need to be treated like software… the same way we have npm and pip for managing dependencies, we need a way to manage skills like that." Not a metaphor — a checklist: vulnerabilities management, documentation, integration, wrappers, code reviews, deployment, CI pipelines, distribution reviews.
- THE SHARPER CLAIM. Skills describe intent, in English, not just capability — which is why they are proposed as the successor to code packages, not a variant of them.
The catalogue — /catalogue/ — build first, per the brief
The registry v0: eight skills, generated from /catalogue/skills.json, never hand-written. Per skill: name, repo, trigger summary, do-NOT-use clause (captured for 1 of 8), and the version/permissions fields the thesis demands that none declare. /catalogue/divergence.html is the exhibit: the two copies of use-sg-playwright, and the gap between them measured at build time — the thesis's predicted failure, live, on the estate's own skills.
The authoring guide — /authoring/ — usable the same day
Five conventions extracted from the eight shipped skills, none of them documented anywhere else: (A) the trigger-rich description, (B) the do-NOT-use clause, (C) self-containedness (embed the schema, don't require a clone), (D) cross-reference canonical contracts rather than copy them, (E) named-trigger vocabulary drawn from the domain's own terms. /authoring/when-not-to.html carries the estate's best governance artefact — the sg.llm skill refusal — generalised into three rules: skills compete for triggering; a skill must not become a second source of truth; skills earn their keep on distinct tasks.
The thesis — /thesis/ — skills are software packages
The lifecycle scorecard: version control (partial), dependency management (none, and the duplicate is the consequence), vulnerability management (one review, never repeated), code review (none), CI (none), distribution (none), documentation (yes). One green cell out of seven. Plus the forking ecosystem — "forking a skill is good; sync back via pull-request-equivalents" — and a five-step honest sequencing from a version field to the graph projection research programme.
The graph — /graph/ — design, not shipped
"Today's skills are static photographs of what they should be." A skill decomposed into typed primitives (principles, concepts, facts, knowledge, actions) across multiple graph types, projected in the context of use. Every shipped skill in the estate is the static photograph this page criticises — stated plainly on the page itself.
Identity and permissions — /identity/ — the most consequential, least built
A skill "must come with the identity and permission set needed to execute it." The PBOM (Permissions Bill of Materials) generalisation to every unit of execution. The OAuth critique: it "forces you to grant a hundred percent when you need ten percent." The proposed frontmatter field (permissions: {network, vault, spend}), declared not enforced. Labelled explicitly as a critique with a direction, not an implementation.
The skill economy — /economy/ — an economic position, not a fact
Base vaults as the unit ("far more than a SKILL.md… all the materials that make the skill explainable, trustworthy, and executable"). The branded/certified/customised cascade — "the customised version… is fundamentally what you sell." Distribution via aggregators (Tessl, named). Scoring: "provenance, evals, and an auditable graph, rather than gameable popularity." One dated external validation (1 June 2026 conference feedback), cited as one signal, not proof.
What is actually shipped — /shipped/ — unsoftened
The scorecard, the duplicate, zero permissions, zero evals, no registry, the February 2026 review precedent that was never repeated, and the static-photograph criticism applied to the estate's own artefacts. /shipped/questions.html publishes six open questions with no answer anywhere in the corpus (what is a skill version when the interpreter changes underneath it; how do you test intent) and five honest tensions.
The network — /network/
wardley-maps.sgit.ai owns the skill lifecycle · open-source.sgit.ai owns the OWASP monetisation thread and fork economics · graphs.sgit.ai owns meaning through connectivity · coding.sgit.ai owns /for-agents/ · standards.sgit.ai owns vault-as-substrate · llms.sgit.ai owns sg.llm.* · sgit.ai is the parent project. Full table at /network/.
Site map
/ · /catalogue/ · /catalogue/divergence.html · /catalogue/skills.json · /authoring/ · /authoring/when-not-to.html · /thesis/ · /graph/ · /identity/ · /economy/ · /shipped/ · /shipped/questions.html · /documents/ (+ 10 readers) · /agents/ · /network/ · /about/participant.html · /admin/comms.html · /admin/versions.html · /admin/index.html · /llms-full.txt · /index.md