Known Limitations

This collection is deliberately opinionated and useful within a specific scope. Being clear about where it stops is part of using it well.

It depends on your project context — by design

This is a deliberate design choice, not a defect: the skills are built to write in your voice, so they need context to do it. Without discovery (purpose.md, buckets.md, and especially tonality.md), the production skills fall back to safe, generic defaults and drift toward a sales-y, everyone’s-voice register. They work without configuration, but they are not meant to produce your voice from nothing. Treat the discovery setup as a prerequisite, not an optional extra.

It is opinionated, not universal

The frameworks, quality criteria, and defaults are closely aligned with one writer’s workflow and a particular school of social-first online writing. They encode real opinions about structure, pacing, and voice. If your conventions differ (academic writing, long-form journalism, technical reference docs), expect friction — the skills will nudge you toward their house style.

American English first; experimental German support

The collection is American English first — most skills, guides, and examples assume American English conventions. German support is experimental: a smaller, intentionally scoped subset — source analysis (analysiere-quelle), long-form drafting (schreibe-entwurf), EN→DE rewriting (rewrite-de), and style (finde-stil).

The German pipeline has known gaps: short-form German drafting (schreibe-post) and German micro-post distillation (destilliere) are referenced as future skills and do not exist yet. See the German workflow for what is currently covered. Other languages are not supported.

It is not a one-shot content generator

The skills package and structure your thinking — they do not replace it. You still have to decide what matters, choose the point of view, supply the raw material, and take responsibility for the result. Hand the skills an empty prompt and you get generic output; hand them real ideas and they help you shape them.

Producing something good is also iterative: expect to move between the discovery, drafting, and review skills more than once, refining input and output as you go. This is not a press-one-button-get-a-finished-piece tool.

No SEO optimization

The workflows are written for human readers (and, incidentally, AI readers). There is currently no support for search-engine optimization — keyword targeting, meta structuring, or SERP-oriented formatting. If you need SEO, you will have to layer it on yourself.

Imagery is Midjourney-only (experimental)

The illustrate skill is experimental and produces Midjourney prompts only. It does not target other image generators, search stock photography, or edit existing images, and the end-to-end image workflow (submission, asset capture, wiring) is only partially built.

Source material handling is evolving

The skills draw on external concepts and may use reference texts during the writing process. How sources are used — and whether they are named in the output — is still being refined. By default the skills use ideas without explicit attribution unless you ask for it. This collection is not a substitute for, endorsement of, or official companion to the works it is inspired by.

Output is non-deterministic

These are LLM-driven skills. The same input can produce different output across runs, and quality varies with the model and runtime you use — features that work in one runtime or model tier may behave differently in another. The evaluation skills (rate, diagnose) are guidance, not guarantees — they reflect the framework’s judgment, not an objective ground truth. And because they are themselves LLM-driven, they are subject to the same non-determinism: a score or diagnosis can vary between runs. Always review output before publishing.

It is token-intensive

This is not a lightweight workflow. The skills load reference guides, your configuration, and your voice samples into context; the discovery skills run multi-turn interviews; and the drafting skills produce long outputs (up to 2,500–3,000 words) often across several iterations. Expect meaningfully higher token usage — and, on metered runtimes, higher cost and slower turns — than a single one-shot prompt. Treat it as a deliberate trade: more context and structure in exchange for output that actually sounds like you.

Configuration lives outside the agent context — by design

The writer-specific configuration is not part of the agent or plugin setup. By design it lives in each project’s .online-writing/ folder, separate from the skills’ own context, so one installation can serve many projects without mixing their voice and strategy (see Configuration). You are responsible for creating, curating, and keeping it current — including using separate project folders when contexts should not mix. Nothing is synced or stored centrally for you.