Public methodology · open record

How we know a prompt is safe to ship.

Loddle does not run a prompt marketplace. Every prompt in the library passes four quality gates before publication: schema validation, automated citation-detection validation of every sample output, per-niche safety architecture embedded in every copy, and a transparency badge showing real generation metadata. This page describes what each gate actually does.

Quality gates

Schema · generation · safety · transparency

Model

claude-sonnet-4-6

Sample generation model

Prompts

150

Across 4 professional niches

Schema gate

Every build, automatically

§ 01 · The four quality gates

Every prompt clears four gates before it ships.

The same four gates apply to a Basic admin-efficiency prompt and an Advanced equitable-adjustment letter. The difference is the height of the bar at each gate, not the number of gates.

Quality gate · empirical test

Tested with a dedicated generation script — validated on every build.

Every Loddle prompt has a sample output generated by claude-sonnet-4-6 using a niche-specific safety system prompt, produced by a dedicated generation script — not automatically on every build. Outputs that contain professional codes (CPT, ICD-10, FRCP, ASTM, and similar) are checked by an automated citation-detection validator on every build; where automated regeneration was not possible, the affected sample was manually redacted to a placeholder pending regeneration.

Sample generation · dedicated script; citation check · every build

Model

·claude-sonnet-4-6

Citation detection

Outputs containing professional code citations (CPT, ICD-10, FRCP, ASTM, etc.) are checked automatically on every build. Where automated regeneration was not possible, the affected sample was manually redacted to a placeholder.

Scope

Every prompt in the library — all 150 across all four niches — goes through the same generation pipeline.

Transparency

The model name and generation date are stored with each sample and displayed on every prompt page via the transparency badge.

Quality gate · contract & format

Structured to a validated schema.

Every prompt is validated against a JSON schema enforced by the prebuild CI step. The schema requires role assignment, task definition, a constraints array, an uncertainty instruction, a disclaimer, and a minimum number of FAQ entries. A prompt that fails schema validation blocks the build — it cannot reach the library.

Schema validation · every build, automatically

Required fields

·Role — who the AI is being asked to play
·Task — what the user wants out
·Constraints — what the AI must not do
·Uncertainty instruction
·Professional disclaimer
·FAQ entries

Enforcement

JSON schema validation in the prebuild script. A prompt cannot reach the library without passing schema validation.

Gate

The validate-prompts.ts script runs automatically before every build via the npm prebuild lifecycle hook.

Format

Plain text. No model-specific syntax. Output is portable across model providers.

Quality gate · domain safety

Per-niche safety architecture baked in — and embedded in what you copy.

Every prompt carries constraints written for its specific profession: a lawyer's billing-narrative prompt knows what a contingency arrangement is; a contractor's RFI prompt knows what a CCD is; a healthcare-admin patient-comms prompt knows what HIPAA permits in a status email. Those constraints, the uncertainty instruction, and the professional disclaimer are not just displayed — they are embedded in the text you copy, so the model always receives them regardless of where you paste it. Runtime injection during in-product execution is planned for the Prompt Runner beta — join the waitlist to be notified when it opens.

Niche constraints · applied at copy today, runtime planned

Scope locks

Every prompt declares its profession scope. Cross-domain use is flagged on the prompt page.

Embedded constraints

Safety constraints, uncertainty instruction, and disclaimer are prepended to the prompt text you copy — the model always receives them regardless of where you paste the prompt.

Runtime injection

The Prompt Runner (in-product execution) is not live yet — it is a planned beta. Join the waitlist to be notified when it opens.

Out-of-scope flag

If your matter falls outside the prompt's declared scope, the prompt page tells you so before you copy.

Quality gate · transparency

Confidence tiers reflect the stakes — with honest disclaimers.

Confidence tiers — Basic, Professional, Advanced — reflect the stakes of the task and are displayed on every prompt. Every prompt carries a profession-appropriate disclaimer; the language is consistent across the library to avoid drift. The transparency badge on each prompt page shows the model used, the generation date, and the confidence tier — all drawn from real data.

Confidence tier · displayed on every prompt

Confidence tiers

·Basic — routine, lower-stakes administrative tasks
·Professional — practice-grade tasks; review before use
·Advanced — high-stakes; use with professional review

Disclaimer

Each prompt carries a profession-appropriate disclaimer; the language is fixed across the library to avoid drift.

Transparency badge

Shows the model name (claude-sonnet-4-6), generation date, and confidence tier — sourced from the prompt data, not authored by hand.

Liability

Loddle's terms make clear: every output is informational; final professional judgement is required for every deliverable.

§ 02 · The pipeline

From authored prompt to published page.

Every prompt follows the same pipeline. Schema validation blocks the build on any failure; sample generation is run via a dedicated script, and its output is re-validated automatically on every build; nothing ships without passing both gates.

Step 1in pipeline

Prompt is authored and curatedDrafted with profession-specific constraints, an uncertainty instruction, and a professional disclaimer. Declared variables are enumerated for the Prompt Runner.

Curator · profession-focused

Step 2in pipeline

Schema validation gateThe validate-prompts.ts prebuild script checks every required field. Any failure blocks the build — the prompt cannot reach the library.

Automated · prebuild CI

Step 3in pipeline

Sample generation (dedicated script)claude-sonnet-4-6 generates a sample output using the niche safety system prompt, run via a dedicated generation script. Outputs containing professional code citations are checked automatically on every build; where automated regeneration was not possible, the sample was manually redacted to a placeholder.

Manual run · generate-samples.ts

Step 4in pipeline

Published with transparency badgeThe prompt ships with its model name, generation date, confidence tier, and embedded constraints visible on the prompt-detail page.

System · CI / publishing

§ 03 · Expert advisors

Practising professionals as named advisors.

We are in the process of engaging practising professionals in each niche as expert advisors; their credentials and review scope will appear on this page and on individual prompt pages when they have been confirmed.

Expert advisors · coming

→

One per niche — lawyers, architects, healthcare admin, contractors.

When confirmed, each advisor's credentials, scope of review, and niche-specific sign-off criteria will be documented here and linked from the individual prompt pages they have reviewed. We will not publish advisor names or credentials until the relationship is confirmed and the scope is agreed.

Status · in progress

Scope (planned)

Vocabulary accuracy, profession-appropriate constraints, disclaimer language, confidence tier assignment.

Credentials

Will be documented on this page when confirmed — not before.

Prompt pages

Prompts reviewed by a confirmed advisor will carry an advisor badge on their detail page.

The library is small because we ship nothing that hasn't cleared every gate.

Loddle editorial standard · adopted v2.0, 2025

Open any prompt. Read the metadata.

Every prompt-detail page shows the model, generation date, confidence tier, and embedded safety constraints. You can read exactly what the AI was told before you decide whether to trust it with client work.

Browse prompts →Get in touch