| name | house-style |
| description | The default writing voice for Dataspheres AI page content — blog posts, reports, docs, landing copy. A measured, white-paper register that reads like a knowledgeable human wrote it, NOT like an AI rushed it. Use this voice by default whenever composing prose for a page; invoke explicitly to revoice existing copy that "sounds like AI". Defines the AI tells to strip out and the register to write in, with before/after examples. |
| argument-hint | (applied by default to page prose) | revoice <page-or-text> | check this draft |
House Style — Write Like a Human, Read Like a White Paper
This is the default voice for all Dataspheres page content: blog posts, research/intelligent reports, documentation, and long-form landing copy. Unless the user asks for a different register (playful, terse marketing, press release), write in this voice.
The target, in the user's words: "white-paper tone without being a white paper." It should read like a knowledgeable person explaining something they understand well — measured, declarative, unhurried — and it must not read like an AI that rushed it.
The register, in one paragraph
Lead with the problem and the domain. Explain mechanisms plainly. Use longer, flowing paragraphs that build an argument through explanation, not through punchy one-liners. State things directly and let the substance carry the weight. Assume an intelligent reader who wants to understand the thing, not be sold the thing. Cite real numbers matter-of-factly. Never announce the structure ("In this post we will…"), never perform ("Here's the thing…"), never say it's a white paper.
Strip these AI tells (this is most of the work)
A draft "sounds like AI" because of a recognizable set of habits. Remove them:
| Tell | Example to kill | Why it reads as AI |
|---|
| Rhetorical triplets | "It shows up in benchmarks. It shows up in pull requests. It shows up in that quiet moment when…" | The rule-of-three cadence is the single most recognizable LLM tic. Say it once, plainly. |
| "Not X, but Y" / "It is not X. It is Y." | "It is not a model-quality problem. It is a process problem." (repeated every few lines) | One per article, maybe. As a reflex, it's a tell. |
| Fragment-for-drama | "There is only hope. And every tool is built on hope." | Sentence fragments deployed for punch. Fold into a real sentence. |
| Clickbait / cutesy headings | "The Spec Wars of 2026" | Use plain, descriptive headings: "Where current tooling stops". |
| Self-announcing transitions | "Here is what separates X from Y:" / "Now, here's the thing:" | Just make the point. "The other structural difference is where state lives." |
| Announcing the subject/thesis | "…is the subject of this post." / "the rest of this post describes one answer to" / "In this piece we explore…" | Don't label the topic — state it and move. "That gap is structural, it is measurable, and no tool closes it for you." Just do it; don't say you're doing it. |
| Em-dash drama | three em-dash asides per paragraph | Em-dashes are fine sparingly. Stacked, they read as generated. |
| Marketing closers | "If you've ever been burned by X, this is worth 5 minutes. The gap is real, it is measurable…" | End on substance, not a CTA pep-talk. |
| Hype intensifiers | "revolutionary", "game-changing", "seamless", "powerful", "robust", "delve", "leverage", "in today's fast-paced world" | Cut or replace with a concrete claim. |
| Competitor dunking | quoting a rival's tagline to sneer at it | State what each tool does, neutrally. Let the comparison speak. |
| Second-person hype | "Imagine an agent that…" as a hook, "you'll love…" | A short concrete scenario is fine; a hype hook is not. |
| Hedge stacking | "It's worth noting that it's generally the case that…" | Say the thing. |
Do this instead
- Open on the problem and the domain, not on a hook. The first job is to make the reader understand what's at stake, plainly.
- Explain the mechanism. When you make a claim ("agents rubber-stamp"), follow with why it happens, in plain cause-and-effect.
- Prefer longer paragraphs that develop one idea fully over staccato one-sentence paragraphs.
- Declarative voice. "The board survives IDE restarts." Not "What's powerful here is that the board survives IDE restarts."
- Numbers, stated flatly, with their source named in the sentence. "AI-generated code contains 1.7× more bugs (code-quality studies, 2025–2026)."
- Plain, descriptive headings that say what the section is about.
- End on the idea, not a pitch. The last paragraph should land the central point, not ask for the click.
- One scenario, concrete, up front is allowed and often good — as long as it's specific, not a hype device.
Before / after (from a real revoice)
Headline
- ✗
The Spec Wars of 2026
- ✓
Where current tooling stops
Triplet → plain
- ✗ "It shows up in benchmark evaluations. It shows up in pull requests. It shows up in that quiet moment when you realize the tests were passing because the agent rewrote them."
- ✓ "In benchmark settings, agents have been observed reaching near-perfect scores without solving the underlying tasks. In ordinary use it appears more quietly: a test suite that passes because the agent adjusted the tests until it did, rather than because the implementation became correct."
"Not X, it's Y" + fragments → measured
- ✗ "The rubber-stamp problem is not a model quality problem. It is a process problem. Without a gate, there is no enforcement. There is only hope. And every SDD tool on the market is built on hope."
- ✓ "The tendency to mark work complete without verifying it is not mainly a question of model quality. It is a property of the process, and it shows up across capable and less-capable models alike. When the only record of completion is the agent's own account, completion and the claim of completion become indistinguishable."
Self-announcing transition → direct
- ✗ "Here is what separates all-dai-sdd from every IDE-based SDD tool: the state lives in Dataspheres AI, not in your editor."
- ✓ "The other structural difference is where the project state lives. In an IDE-based SDD tool, the plan and its status live in the editor. In all-dai-sdd they live on Dataspheres AI."
Marketing closer → substantive close
- ✗ "If you have ever been burned by an AI agent that rubber-stamped a task Done, all-dai-sdd is worth 5 minutes. The verification gap is real, it is measurable, and the enforcement layer is already built."
- ✓ "The verification gap it addresses is not particular to any one tool — it is a property of handing work to an agent and trusting the agent's own account of whether the work got done. Closing it means checking the evidence instead."
Reference revoice (full before/after): dataspheres-ai/specs/sdd-blog-post/republish-v3.mjs, published at dataspheres.ai/pages/dataspheres-ai/spec-driven-development.
Apply, then self-check
After drafting (or revoicing), read the piece once against this list:
Scope and exceptions
- Default for: page bodies, blog posts, research/intelligent reports, docs, long-form landing copy.
- Not this voice: UI microcopy, error messages, button labels, email subject lines (those have their own constraints), or when the user explicitly asks for a different register.
- Email copy keeps its own rule: plain and direct, never exclusivity-hype ("You're in.") — consistent with this voice but governed separately.
- Pairs with the
data-viz skill for diagrams and the rich-content skill for page structure; this skill governs the prose voice within them.
