The engine behind Create

PrayerEngine

Name what is present, receive a fitting practice, and keep agency over what you use.

PrayerEngine is the composition layer inside LumenNous. It turns a plainly stated need into a small, structured practice — assembled from reviewed material, on your device, in the time it takes to press a button.

Why it exists

A blank page is a poor companion when you are anxious, grieving or simply tired. PrayerEngine removes the open-ended task: you say what is present, and it returns language, attention and structure you can use immediately — without asking you to register, subscribe or explain yourself twice.

It is deliberately not an oracle. Selection is bounded and explainable, results are never framed as messages meant for you, and nothing here claims spiritual authority or a guaranteed outcome.

It does

  • Classify a need against a researched taxonomy of intentions.
  • Assemble prayers, affirmations, practices and prompts that were written and reviewed by people.
  • Respect tone, language, length and the things you asked it to avoid.
  • Run locally, offline, in under a second.

It does not

  • Diagnose you, or read a random draw as a sign.
  • Promise healing, protection, wealth or any external result.
  • Replace medical, psychological or crisis care.
  • Keep your words, build a profile or track a streak.

How a request becomes a practice

Seven steps, all of them synchronous and all of them in your browser.

  1. Validate

    The request is parsed against a strict schema before anything else happens: need text, intention, output form, duration, tone, language and any avoidances. Malformed input fails here, not halfway through composition.

  2. Route for safety

    A conservative classifier scans the words for crisis, acute-distress and metaphysical-escalation signals. Anything it flags stops normal composition and returns fixed, human-written support language instead.

  3. Classify the need

    An explicit intention is always honoured. If you choose “not sure”, the need is scored against the research taxonomy — names, aliases, intentions and descriptions — and the closest active category wins.

  4. Filter the library

    Candidates are narrowed by output form and duration, ranked for tone and language preference, and anything matching an avoidance you named is removed outright rather than down-weighted.

  5. Draw without repeating

    Each context keeps its own shuffle bag. Every compatible item is used once before the bag refills, and a refill can never open with the item you just received.

  6. Assemble and re-validate

    Prayer, affirmation, practice steps and a reflection prompt are composed into one recipe, which is validated again before it reaches the screen. Coherence output keeps its reviewed five-stage order and fixed three-minute core.

  7. Remember almost nothing

    Only content IDs, cycle counters and short result fingerprints are written to this browser. The words you typed are never stored, never sent anywhere by the local engine, and never used to build a profile.

What you can steer

Every control is optional, and every one of them changes what the engine is allowed to select. Avoidances are exact: name an image or a word you do not want, and matching material is dropped from the pool entirely.

Form
PrayerAffirmationMeditationCoherence practice
Length
Brief · 2 min5 min10 minExtended · 15 min
Tone
GentleDirectContemplativeDevotionalGrounding
Language
CreatorSourceDivineGnostic termsNeutral

Variety without randomness theatre

Familiar structures help; identical results do not. Each combination of intention, form, length, tone and language keeps its own bag of compatible items. The engine draws from that bag until it is empty, then reshuffles — and never opens a new cycle with the item you saw last.

Once each

Every compatible item is offered before any repeats.

No seam repeat

A refilled bag cannot begin with the previous draw.

Bounded

Cycling is a variety mechanism, never a message or a sign.

Safety comes before composition

The classifier runs before any content is drawn, and it is deliberately conservative: a prayer for courage before an exam is not a crisis. When it does flag something, the engine stops and hands over to fixed wording written and reviewed by people.

  • Crisis

    Self-harm, suicide, abuse, medical emergency or acute psychosis signals return immediate encouragement toward trusted people and professional help. No assembled content is shown.

  • Escalation risk

    Language about attack, possession or contamination is met with calm grounding and an explicit refusal to confirm supernatural harm — never with deliverance framing.

  • Distress

    Heavier emotional wording receives a grounded pause and a gentle pointer toward support, rather than a longer practice.

Matched phrases are used in-process only. They are never logged, never persisted and never sent anywhere.

What it draws from

Nothing is invented at request time. The engine only arranges material that already exists in the reviewed library, where every item carries its classification — historical text, modern interpretation, practitioner tradition, research or LumenNous editorial — and its sources.

active intentions
26
reviewed prayers
72
affirmations
96
guided practices
47
reflection prompts
140
coherence sessions
5

Content validation runs at import time. Duplicate IDs, broken references, missing category coverage and unsupported evidence labels fail the build rather than reaching a screen.

Where your words go

Nowhere, for the local engine. Composition happens in the page you already loaded — there is no account, no database and no analytics behind it.

  • Request text is never written to storage.
  • History holds content IDs, cycle counters and short fingerprints only, in this browser.
  • You can clear that history at any time from Saved.
  • Third-party audio loads only after you explicitly ask for it.

The one exception

In this test build, choosing to describe a need in your own words — rather than picking a specific intention — sends that request through a server-side AI boundary for a bounded original composition. The key stays on the server, the route is disabled unless it is explicitly enabled, and fixed safety responses always bypass it. Choose an intention and the flow stays entirely local.

Where the development stands

  • Classification, filtering, cycling, safety routing and assembly are implemented and unit-tested.
  • The three-minute embodied coherence method is versioned, with five daily session variants.
  • Generated sessions can be copied or shared, but are not yet saved as complete custom recipes.
  • Research snapshots are promoted manually; some seed material is still marked as draft pending final editorial review.

Start with what is actually present

No account, no feed, no streak. Name the need and see what the engine returns.

More on the method in About and on data in Privacy.