Why Chat Boxes Do Not Compose
The first wave of AI in regulatory writing was the chat box. A writer pastes a section, asks a model to tighten it, copies the result back. It works for individuals. It does not compose, it does not version, and it leaves no trail a quality assurance lead or an inspector can read.
A regulatory document is not a prose problem. It is a compliance problem dressed up as a prose problem. The standards that govern a clinical study report — ICH E3, the FDA Common Technical Document guidance, the agency-specific reviewer expectations — are public, versioned, and dense. They do not change when you switch from GPT to Claude. They change when a regulator publishes a new guidance, or when a senior medical writer ratifies a different interpretation. A chat box treats every interaction as a fresh string. A registry treats every interaction as the invocation of a named, versioned standard.
The Skill Registry, in One Sentence
A skill in Asthra is a manifest. It declares what it does, which document types it applies to, which standards it references, and what shape of output it returns. The same manifest powers the slash menu in the editor, the chips a writer can click, and the chat agent's tool calls. There is one source of truth across the entire surface area. When a subject matter expert authors a new rubric — for an FDA PSUR, an EMA periodic benefit-risk evaluation, a device clinical evaluation report — it ships as a manifest, not a code change. Network effects compound.
What an Audit Trail by Construction Looks Like
Every skill invocation produces a structured record. Skill identifier, version, rubric references, verdict, findings. The findings are typed, not free text. A persona reviewer that flags a missing benefit-risk statement returns a finding object with the rubric anchor — ICH E2C(R2), section 18.3 — alongside the prose snippet that triggered it.
That record is fit for quality assurance, fit for inspection, and fit for the kind of regulator-facing API submission that is on the horizon. The trail exists because the registry forced it into existence, not because someone remembered to write it down.
Six Skill Kinds, Each With a Reason to Exist
Persona skills represent reviewer perspectives. An FDA reviewer skill applies the rubric an FDA reviewer would apply, and returns findings in the shape a writer can act on.
Chat-prompt skills are conversational moves a writer can take inside the agent, the kind of prompt that used to live in a personal cheat sheet.
Style skills are prose-level transformations — make this more concise, make this more declarative — pinned to house style.
Validation skills run deterministic checks: cross-reference integrity, defined terms used before definition, statistical claims with no upstream source.
Surgical-edit skills are typography-grade operations on the document, the kind that should never go through a paragraph rewrite.
Composite skills chain the others into named workflows that ship as one button.
Doctype-Aware by Design
A skill declares the document types it applies to. A CSR-only skill does not surface for a CMC writer. A device CER skill does not pollute a drug PSUR menu. This is small, and it is the difference between a copilot that scales across the regulatory portfolio and a copilot that feels noisy in every individual role.
Progressive Disclosure of Capability
Writers do not have to know what is possible. They type a slash, or click a chip, and the registry surfaces what applies to the document they have open. When a writer pastes a claim, the chat agent suggests the validation skill that would check it. The writer learns the standard by using the tool.
The catalogue itself has a progressive-disclosure path planned — once the registry passes roughly forty skills we expect to surface them hierarchically rather than as a flat list — but the trigger is shipped capability, not aesthetic preference.
What This Unlocks
The registry is the layer that makes everything else in Asthra defensible. The hyperlinking work we shipped this past week lives on top of it. The literature search that now happens inside the chat lives on top of it. The persona reviewers, the validation passes, the surgical edits — all of them live on top of it. Each skill is a contract a regulator could in principle inspect. Each invocation is a row in an audit table.
The Veeva-shape moat in regulatory technology has always been the system that knows what was done, by whom, against what version of which standard. We are building that system in the layer where the work actually happens — inside the document.