"Prompt engineering, system prompts, document-skills, and SCIENTIA (research 2026)"

Prompt engineering, system prompts, document-skills, and SCIENTIA

This page records research findings on prompt engineering and system-prompt design, and maps them onto Vox systems: continuation prompts, ARS skills, documentation extraction, and SCIENTIA publication flows.

It is research guidance, not a shipped contract. Contract and policy surfaces remain in contracts/, CI gates, and crate-level SSOT documentation.

Executive summary

  1. Prompt quality depends more on layered instruction architecture than on one large prompt.
  2. Skills-as-documents is now an industry-standard pattern; Vox can reuse this pattern with existing ARS trust and sandbox controls.
  3. Document ingestion and retrieval increase indirect prompt-injection risk and require explicit trust boundaries.
  4. SCIENTIA automation must preserve human accountability for claims, ethics, and venue disclosures.
  5. Legacy submission ecosystems (journal portals, arXiv workflows, DOI metadata channels) require explicit AI-use disclosure and citation integrity checks.

What external guidance converges on

Layered instruction design

Long-context behavior and recency

Long-context studies and vendor practice show strong positional bias in model attention. In practical terms, this supports keeping durable policy short and relocating session-critical behavioral reinforcement near the active context edge (for example continuation prompts and machine-verifiable gates).

References: Lost-in-the-middle summary, Found in the Middle paper index, arXiv:2406.02536.

Skills-as-documents and progressive disclosure

External ecosystems now package reusable agent capabilities as markdown plus front matter:

This aligns with Vox SKILL.md concepts documented in Vox Skill Marketplace. It also aligns with ARS support for SkillKind::Document and trust-aware runtime policies in vox-skills.

Prompt security and untrusted document flows

Threat model

Implication for Vox document workflows

When using skills, docs, or publication metadata as context, default posture should be:

  • trusted instructions are explicit, versioned, and bounded,
  • retrieved documents are treated as untrusted data until validated,
  • policy and quality gates remain outside model free-form output.

SCIENTIA and legacy publication implications

SCIENTIA publication automation already encodes hard boundaries for fabricated or undisclosed AI use in SCIENTIA publication automation SSOT and companion publication readiness docs.

External publication policy direction is consistent:

Policy sourcePractical implication for Vox SCIENTIA
COPE AI tools positionAI cannot be an author; humans remain accountable.
ICMJE AI use by authorsDisclosure in submission workflow and manuscript body is expected.
WAME revised recommendationsTool/version/method disclosure and author responsibility.
Nature AI policyDisclosure requirements and stricter controls on generated media.
Elsevier journal AI policyMandatory disclosure and human verification of references/claims.
arXiv AI tool policySignificant AI use disclosure; authors own all content quality.
IEEE AI text guidanceDisclosure in article sections and strict accountability.
BMJ AI use policyNatural-person authorship and explicit usage disclosure.
JAMA reporting guidanceStructured reporting of tool details and usage surface.
Crossref metadata requirementsMetadata completeness and provenance remain mandatory.
Zenodo software metadata guidanceDeposit metadata integrity (CITATION.cff, .zenodo.json) is operationally important.

Legacy systems

Legacy systems in this context means journal web portals, email-driven editorial pipelines, and manually mediated archive submissions. These systems still require human attestation, policy-aware disclosures, and rigorous citation checks. Prompt libraries and document-skills can accelerate preparation, but cannot replace accountable authorship workflows.

Integration guidance for Vox

flowchart TB
  subgraph instructionLayers [InstructionLayers]
    agentsRules[AGENTS_md_And_Overlays]
    continuationPrompt[ContinuationPrompt]
    arsSkills[ARSSkills_DocumentKind]
    docsCorpus[DocsFrontmatter_And_Body]
  end
  subgraph enforcementLayers [EnforcementLayers]
    ciGates[CIAndTOESTUB]
    socrates[SocratesEvidenceAndRisk]
    preflight[PublicationPreflightAndWorthiness]
  end
  instructionLayers --> modelOutput[ModelOutput]
  modelOutput --> enforcementLayers
  docsCorpus --> mensPairs[MensDocsPairs]

Near-term, low-risk moves

  1. Publish venue-specific document-skills (for disclosure templates, checklist transforms, and metadata hygiene) using existing ARS trust boundaries.
  2. Keep policy gates deterministic and machine-checkable (publication_preflight, Socrates evidence checks, CI contracts).
  3. Add explicit disclosure fields in publication metadata pathways where needed, while preserving current SSOT ownership.

Research-to-implementation boundaries

  • Do not treat citation or readership projections as hard publish gates by default.
  • Do not allow free-form model outputs to bypass digest-bound approvals or preflight findings.
  • Do not mark policy claims as shipped until linked code paths and contracts exist.

Bibliography (external)