Prompt Commons: Community Prompt Governance

• Prompt Commons is a community-managed repository of LLM prompts that integrates governance metadata, versioning, and civic oversight.
• It employs open, curated, and veto-enabled protocols to ensure diverse value representation and rapid remediation of harmful outputs.
• Empirical evaluations show that collective management enhances neutrality and stakeholder satisfaction while significantly reducing response times to harmful content.

Prompt Commons designates a versioned, community-maintained repository of prompts for LLMs, structurally integrating governance metadata, licensing, and layered moderation to treat prompt engineering as a first-class civic artifact. Its objective is explicit alignment of AI outputs with pluralistic and local value judgments in urban governance settings, addressing the prompt-sensitivity of LLM behavior and the risk of dominant value capture by isolated prompt authors (Mushkani, 15 Sep 2025). The framework recognizes prompts as loci of normative influence and establishes procedures for their transparent, contestable, and plural management.

1. Foundational Concept and Rationale

Prompt Commons is predicated on the observation that LLMs exhibit extreme prompt-sensitivity: minimal changes in prompt phrasing can induce substantial shifts in tone, policy position, inclusivity, and underlying value judgments. In urban domains such as public-space design or municipal policy, this sensitivity entails that individual prompt authors' implicit values can become the de facto default if prompt governance is opaque or individualized. Prompt Commons reframes prompts as civic commons artifacts—rendering value judgments explicit, subject to plural input, version-controlled, and auditable before model deployment. The repository is annotated with metadata including author group(s), locale tags, explicit value claims, semantic versions, and change logs, supporting traceable lineage and collective oversight (Mushkani, 15 Sep 2025).

2. Community Dataset and Augmentation Protocols

The core Prompt Commons corpus was constructed from 443 human-authored prompts solicited from a diverse cross-section of civil-society partners in Montréal, representing disability advocates, seniors, immigrant communities, LGBTQ+ organizations, and urban practitioners. Each prompt encodes a scenario plus declared value priorities (e.g., accessibility, biodiversity, transit trade-offs). Coverage and lexical diversity were expanded by applying duplicate removal, paraphrasing, and scenario expansion, yielding an augmented corpus of 3,317 prompts—a 7.5× increase. Descriptive metrics include a mean prompt length rising from 22.6 to 31.7 words and vocabulary entropy increasing from 7.53 to 8.39 bits, indicating enhanced lexical variety. Analysis confirms non-trivial representation of equity-related tokens (“wheelchair” in 7.6%, “metro” in 7.3%, “biodiversity” in 4.7%), though groups such as LGBTQ+ (~1.0%) and Indigenous communities (~1.4%) remain under-represented, signalling focal points for subsequent recruitment (Mushkani, 15 Sep 2025).

3. Governance Architectures: Open, Curated, and Veto-Enabled

Prompt Commons operationalizes three distinct governance states to modulate the rigor and inclusivity of prompt contributions:

• Open: Authenticated contributors may propose prompts subject only to lightweight spam, formatting, and basic safety reviews. Prompts are auto-merged following minimal automated scrutiny.
• Curated: Contributions require comprehensive metadata declaration (locale, value claim), and must meet public quotas by author group and topic. Maintainers vet pull requests using a checklist (including controlled-vocabulary value claim, accessibility tags, counter-prompts, and licensing). Approval and merging occur post-discussion on a public audit log.
• Veto-Enabled: The curated protocol is extended: authorized local-stakeholder organizations may issue a minority veto on prompts considered potentially harmful. Flagged prompts are quarantined for a 72-hour appeal and dispute resolution process; upheld vetoes result in prompt shelving.

Each governance regime operationalizes a different balance between openness and collective oversight, with the ability to adapt to evolving institutional or civic priorities (Mushkani, 15 Sep 2025).

4. Empirical Evaluation: Neutrality, Satisfaction, and Remediation

Benchmarks assess the influence of governance state on LLM output, using a contested-choice dataset in which the model must choose one among Left-leaning, Right-leaning, or Neutral/Compromise stances per prompt. Neutrality is quantified as

$$\text{Neutrality score} = \frac{N_\text{neutral}}{N_\text{total}}$$

and decisiveness as

$D = 1 - \text{Neutrality}.$

Key results:

• Single-author prompt (M0): Neutral outcomes in 24% of cases (Left/Right 38% each; D=0.76).
• Commons-governed prompts (M1–M3): Neutrality rises to 48–52% (Left/Right 24–26% each; D≈0.48–0.52).
• Ensemble (M4): Prompt sampling across groups, with explicit compromise instruction, achieves D≈0.49.

Subgroup satisfaction, measured via a 7-point Likert scale across six self-identified stakeholder groups, increased from 4.35 ± 0.86 (M0) to 4.92 ± 0.44 (curated) and 5.48 ± 0.66 (veto-enabled). Satisfaction disparity (Gini coefficient) dropped from 0.096 (M0) to 0.043 (curated).

Remediation speed for harmful outputs was quantified via synthetic incident logs. Across 50 simulated incidents per regime, mean time-to-remediation (±SD) was 30.5 ± 8.9 h (Open), 11.8 ± 3.2 h (Curated), and 5.6 ± 1.5 h (Veto-Enabled), drawn from exponential distributions parameterized by regime-specific means. These statistics demonstrate that collective and veto-enabled processes yield both greater neutrality and faster harm mitigation (Mushkani, 15 Sep 2025).

5. Metadata Structure, Licensing, and Moderation Procedures

Every prompt is accompanied by a standardized metadata schema comprising author group(s), locale tags, claimed values, semantic versioning, and change logs referencing all related discussion or proposals. This infrastructure ensures forensic auditability and traceability to municipal instruments or policy context. Licensing defaults to Creative Commons CC BY 4.0, with optional share-alike (CC BY-SA 4.0); downstream artefacts (such as models or tools) may introduce OpenRAIL-style license conditions (e.g., explicit prohibitions against biometric surveillance or harassment) as contextual constraints.

The moderation protocol consists of three tiers: (i) automated checks (toxicity, personally identifiable information), (ii) checklist-based maintainer review, and (iii) community veto in veto-enabled deployments. This layered system creates explicit defense mechanisms against prompt-set “dominance capture” by any single interest group (Mushkani, 15 Sep 2025).

6. Connections to Reusable Prompting and Data-Efficient Model Adaptation

Prompt Commons aligns with advances in reusable and contextual prompting as described in Super-Prompting strategies (Rezaei et al., 2022), where model-agnostic prompt fragments (concept words, scenario elements, affect cues) are compiled into repositories. These prompt fragments can dramatically reduce annotation requirements for downstream tasks—e.g., achieving full-data performance with only 35–40% of human-labeled data in visual commonsense reasoning contexts. The incorporation of reusable prompt templates and empirical documentation into Prompt Commons enables annotation-efficient research, cross-institutional prompt sharing, and systematic empirical benchmarking of prompt efficacy and transferability (Rezaei et al., 2022).

7. Implications, Extensions, and Future Directions

Prompt Commons operationalizes prompt engineering as a deliberate governance vector, enabling procedural pluralism in the configuration of AI-mediated civic decision support. In urban contexts, the framework addresses the risk of one-sided value imposition by rendering prompt deliberation and value contestation explicit and accountable. Ongoing recruitment of under-represented stakeholders is required to ensure genuinely plural representation in the prompt base. Future research may develop automated tools for prompt selection, ensemble aggregation for scenario coverage, and dynamic license enforcement linked to context-specific policy requirements. The Prompt Commons model provides a scalable paradigm for integrating technical transparency, civic auditability, and plural value alignment in sociotechnical AI deployment (Mushkani, 15 Sep 2025, Rezaei et al., 2022).