Evaluate the 24 Content Creator-scope skills in Varnan-Tech/opendirectory, where the AV half carries a production and vendor cost the writing half does not #1077

New issue

Open

opened 2026-08-15 23:46:57 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-15 23:46:57 +00:00

Member

Source

Upstream: https://github.com/Varnan-Tech/opendirectory, directory skills/. MIT, org-owned, created 2026-04-13, last push 2026-08-14, 606 stars, 64 forks, 1 open issue, roughly 212 commits on main. The contents API confirms exactly 64 skill directories.

Scope

24 of the 64 fall in Content Creator conceptual scope. Exec triage sorted all 64 into tiers against portfolio fit, and this issue carries the Content Creator subset of tiers A, B, and C.

Content Creator owns boundary-suggest-human-comms, so the adoption call on every skill below belongs to that role, not to exec. This issue scopes the candidate set and records the cost axis. It does not decide.

The finding that matters

The 24 split cleanly on adoption cost, and the split does not follow the tier.

Writing cluster - prompt-only - a SKILL.md plus a references/ directory, no runtime dependency, cheap to read and cheap to reauthor
AV cluster - production-coupled - image, video, deck, and GIF generation all imply a rendering toolchain or a hosted vendor that this estate does not run today

Tier C looked low-value on titles alone, but that was fit-to-Kai, not quality. The reason the AV cluster should still be evaluated last is the toolchain cost, not the writing.

Candidate set

Writing and social, prompt-only, evaluate first

show-hn-writer - tier A - launch-post authoring for the channel that actually reaches an infra audience
github-discussion-to-devrel-content - tier A - converts existing repo activity into content, which is the exact pipeline that does not exist today
human-tone - tier B - voice work, read specifically for the delta against voice-corpus and writing-voice-adaptation
noise2blog - tier B - unstructured work notes into a post
cook-the-blog - tier B - blog production pipeline
tweet-thread-from-blog - tier B - repurposing into X and Bluesky
newsletter-digest - tier B - no list exists today, evaluate as latent
email-newsletter - tier B - same
podcast-transcript-fetcher - tier B - source-material acquisition rather than authoring
linkedin-post-generator - tier C - read in full already, see the blocker section

Visual and AV, production-coupled, evaluate second

blog-cover-image-cli - tier B - the only AV item with an immediate concrete use
noise-to-linkedin-carousel - tier C
graphic-chart - tier C - overlaps the existing dataviz skill, which is stronger
graphic-slide-deck - tier C
graphic-case-study - tier C
graphic-ebook - tier C
graphic-gif - tier C
vid-product-launch - tier C
vid-sizzle-reel - tier C
vid-motion-graphics - tier C
hyperframes-product-launch-video - tier C - the name implies a vendor named Hyperframes, unverified

Shared with exec scope, tracked in both issues

oss-launch-kit - tier A - launch timing is an exec call, the announcement content is Content Creator
producthunt-launch-kit - tier B - same split
position-me - tier B - positioning sits across strategy and messaging

What was actually verified

Two of the 64 SKILL.md files were read in full, linkedin-post-generator and claude-md-generator, plus one .env.example and the repo README. Everything else here is title-level and repo-metadata level. Treat the tiering as a sorting hypothesis, not a verdict.

Quality on the two read was better than the marketing framing suggests. linkedin-post-generator carries an eight-step workflow, a roughly 40-word banned list, five format templates, hook formulas, and a references/ split.

Known blockers in the two that were read

linkedin-post-generator reads COMPOSIO_API_KEY and calls linkedin_create_linkedin_post against an OAuth-connected LinkedIn account. That is live publish authority on a human-facing surface, brokered by a third party. It cannot be adopted in that shape.
claude-md-generator shells out to generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent with $GEMINI_API_KEY, which is a backend-direct model call and violates the Agent Proxy transport rule.

Both of the two files opened carried an undeclared external dependency. Inference, not established: the same is likely true across the other 62. The observation that would settle it is a grep for API_KEY, curl, and https:// across all 64 SKILL.md files, which is the first step below.

Acceptance criteria

Grep all 64 upstream SKILL.md files for credential reads, shell-outs, and network fetches, and record which of the 24 below are implicated.
Open each of the 24 in full, including references/ and any .env.example.
Mark each adopt, reauthor, or discard, with a one-line reason.
Name the exact network, credential, or vendor surface for any skill that carries one.
Land accepted skills under .agents/composed/ or .agents/skills/ per the placement rule in tooling-skill-authoring, one issue one commit.
Use no upstream install path. No npx "@opendirectory.dev/skills", no /plugin marketplace add. Both write outside the canonical agent-compose and mcporter projection. Patterns are reauthored locally under MIT attribution.

Companion issue covers the exec-scope subset of the same upstream repo.

## Source Upstream: https://github.com/Varnan-Tech/opendirectory, directory `skills/`. MIT, org-owned, created 2026-04-13, last push 2026-08-14, 606 stars, 64 forks, 1 open issue, roughly 212 commits on main. The contents API confirms exactly 64 skill directories. ## Scope 24 of the 64 fall in Content Creator conceptual scope. Exec triage sorted all 64 into tiers against portfolio fit, and this issue carries the Content Creator subset of tiers A, B, and C. Content Creator owns `boundary-suggest-human-comms`, so the adoption call on every skill below belongs to that role, not to exec. This issue scopes the candidate set and records the cost axis. It does not decide. ## The finding that matters The 24 split cleanly on adoption cost, and the split does not follow the tier. * **Writing cluster** - prompt-only - a SKILL.md plus a `references/` directory, no runtime dependency, cheap to read and cheap to reauthor * **AV cluster** - production-coupled - image, video, deck, and GIF generation all imply a rendering toolchain or a hosted vendor that this estate does not run today Tier C looked low-value on titles alone, but that was fit-to-Kai, not quality. The reason the AV cluster should still be evaluated last is the toolchain cost, not the writing. ## Candidate set **Writing and social, prompt-only, evaluate first** * `show-hn-writer` - tier A - launch-post authoring for the channel that actually reaches an infra audience * `github-discussion-to-devrel-content` - tier A - converts existing repo activity into content, which is the exact pipeline that does not exist today * `human-tone` - tier B - voice work, read specifically for the delta against `voice-corpus` and `writing-voice-adaptation` * `noise2blog` - tier B - unstructured work notes into a post * `cook-the-blog` - tier B - blog production pipeline * `tweet-thread-from-blog` - tier B - repurposing into X and Bluesky * `newsletter-digest` - tier B - no list exists today, evaluate as latent * `email-newsletter` - tier B - same * `podcast-transcript-fetcher` - tier B - source-material acquisition rather than authoring * `linkedin-post-generator` - tier C - read in full already, see the blocker section **Visual and AV, production-coupled, evaluate second** * `blog-cover-image-cli` - tier B - the only AV item with an immediate concrete use * `noise-to-linkedin-carousel` - tier C * `graphic-chart` - tier C - overlaps the existing `dataviz` skill, which is stronger * `graphic-slide-deck` - tier C * `graphic-case-study` - tier C * `graphic-ebook` - tier C * `graphic-gif` - tier C * `vid-product-launch` - tier C * `vid-sizzle-reel` - tier C * `vid-motion-graphics` - tier C * `hyperframes-product-launch-video` - tier C - the name implies a vendor named Hyperframes, unverified **Shared with exec scope, tracked in both issues** * `oss-launch-kit` - tier A - launch timing is an exec call, the announcement content is Content Creator * `producthunt-launch-kit` - tier B - same split * `position-me` - tier B - positioning sits across strategy and messaging ## What was actually verified Two of the 64 SKILL.md files were read in full, `linkedin-post-generator` and `claude-md-generator`, plus one `.env.example` and the repo README. Everything else here is title-level and repo-metadata level. Treat the tiering as a sorting hypothesis, not a verdict. Quality on the two read was better than the marketing framing suggests. `linkedin-post-generator` carries an eight-step workflow, a roughly 40-word banned list, five format templates, hook formulas, and a `references/` split. ## Known blockers in the two that were read * `linkedin-post-generator` reads `COMPOSIO_API_KEY` and calls `linkedin_create_linkedin_post` against an OAuth-connected LinkedIn account. That is live publish authority on a human-facing surface, brokered by a third party. It cannot be adopted in that shape. * `claude-md-generator` shells out to `generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent` with `$GEMINI_API_KEY`, which is a backend-direct model call and violates the Agent Proxy transport rule. Both of the two files opened carried an undeclared external dependency. **Inference, not established:** the same is likely true across the other 62. The observation that would settle it is a grep for `API_KEY`, `curl`, and `https://` across all 64 SKILL.md files, which is the first step below. ## Acceptance criteria 1. Grep all 64 upstream SKILL.md files for credential reads, shell-outs, and network fetches, and record which of the 24 below are implicated. 2. Open each of the 24 in full, including `references/` and any `.env.example`. 3. Mark each adopt, reauthor, or discard, with a one-line reason. 4. Name the exact network, credential, or vendor surface for any skill that carries one. 5. Land accepted skills under `.agents/composed/` or `.agents/skills/` per the placement rule in `tooling-skill-authoring`, one issue one commit. 6. Use no upstream install path. No `npx "@opendirectory.dev/skills"`, no `/plugin marketplace add`. Both write outside the canonical agent-compose and mcporter projection. Patterns are reauthored locally under MIT attribution. ## Related Companion issue covers the exec-scope subset of the same upstream repo.

coilyco-ops referenced this issue

2026-08-16 06:53:08 +00:00

Evaluate the Content Creator-scope skills in AgriciDaniel/claude-seo, where the vendor coupling is declared optional and is actually load-bearing in 16 of 25 #1099

coilyco-ops added the

labels

2026-08-17 02:36:39 +00:00