community-rule/docs/guides/template-recommendation-matrix.md

# Recommendation Matrix — Implementation Context (CR-88)

**Status:** Implemented (CR-88). This doc remains the spec — keep code in sync with it.
**Linear:** [CR-88](https://linear.app/community-rule/issue/CR-88/backend-template-recommendation-matrix-xlsx-sheets-ingestion)
**Roadmap:** [`docs/guides/backend-roadmap.md`](backend-roadmap.md) §4 (`RuleTemplate`) and §13.
**Spec ticket:** [`docs/guides/backend-linear-tickets.md`](backend-linear-tickets.md) Ticket 16.

This doc documents the **method facet matrix** that powers two ranking
surfaces, both consuming the same underlying data:

1. **Create-flow card ranking** — the four card-deck wizard steps
   (`communication-methods`, `membership-methods`, `decision-approaches`,
   `conflict-management`) reorder their `methods[]` array based on which
   methods match the user's selected community facets.
2. **Template grid ranking** — the curated `RuleTemplate` rows shown on
   the marketing home `MarketingRuleStackSection.tsx` and `templates/`
   page get scored by how many of their composed methods match the user's
   facets, then sorted highest-first.

> **Scope note:** Card / modal copy lives in
> `messages/en/create/customRule/*.json` as flat `methods` arrays (one
> entry per method, with `id`, `label`, `supportText`, `recommended`,
> and a `sections` map). The matrix layer adds **facet data** — which
> methods match which user facets — as committed JSON in
> `data/create/customRule/*.json`, validated by Zod and seeded into
> `MethodFacet`. There is no spreadsheet importer, no `xlsx` runtime
> dependency, and the spreadsheets in `~/Downloads/` were a **one-time
> authoring artifact** — they are not part of the runtime contract.

---

## 1. Where things live (post-reorg)

### 1a. Card / modal copy — `messages/en/create/customRule/<section>.json`

Source of truth for all displayed text. Each file holds the page chrome
(`page`, `confirmModal`, `addPlatform`/`addApproach`, `sectionHeadings`,
plus decision-approaches' `sidebar` / `messageBox` / `cardStack` /
`scopeAddButtonLabel`) plus a flat `methods` array. Read via
`useMessages().create.customRule.<section>.methods`. Never duplicated
anywhere else; the recommendation API never returns copy.

### 1b. Facet data — `data/create/customRule/<section>.json`

Mirrors the messages path one-to-one (same filename, different root) so a
content reviewer can read the two side-by-side. Each file is a typed JSON
object mapping each method `id` → its facet matches across the four
canonical groups (`size`, `orgType`, `scale`, `maturity`). Validated by
the Zod schema at seed time and in CI (see §3 — no app-boot validator).
Lives **outside** `messages/<locale>/` because facets are not localized
— they describe the methods, not the UI text.

### 1c. Messages folder reorg (prerequisite)

Today every `messages/en/create/*.json` file sits flat in one folder.
Plan: regroup into the three Figma stages from
[`docs/create-flow.md`](../create-flow.md) §"Product stages":

| New folder | Files |
| --- | --- |
| `messages/en/create/community/` | `informational.json`, `communityName.json`, `communityStructure.json`, `communityContext.json`, `communitySize.json`, `communityUpload.json`, `communitySave.json`, `review.json` |
| `messages/en/create/customRule/` | `coreValues.json`, `communication.json`, `membership.json`, `decisionApproaches.json`, `conflictManagement.json` |
| `messages/en/create/reviewAndComplete/` | `confirmStakeholders.json`, `finalReview.json`, `completed.json`, `publish.json` |
| `messages/en/create/` (root, cross-cutting) | `footer.json`, `topNav.json`, `draftHydration.json`, `templateReview.json`, plus layout-shell strings (`select.json`, `text.json`, `upload.json`) |

Touchpoints: every file move, the imports in
[`messages/en/index.ts`](../../messages/en/index.ts), the namespace shape
exposed via `useMessages()`, and every screen that reads
`m.create.<section>` becomes `m.create.<stage>.<section>`. This is a
mechanical refactor — no behavior change.

**Sequencing (explicit).** This reorg is **its own ticket** and **must
land before** any of §6–§9 below. CR-88's facet JSON paths
(`data/create/customRule/<section>.json`) and `useMessages()` namespaces
(`m.create.customRule.<section>`) assume the reorg is already in place.
Concretely, ship in this order:

1. **Reorg PR (separate ticket).** Move every `messages/en/create/*.json`
   into the table above, update `messages/en/index.ts`, update every
   `useMessages().create.<section>` callsite to
   `useMessages().create.<stage>.<section>`, run `npx tsc --noEmit` and
   `npx vitest run` green. No new behavior.
2. **CR-88 PR (this doc).** Adds `data/create/customRule/`,
   `MethodFacet`, the seed step, and the two API endpoints — all reading
   the post-reorg paths.

If the reorg slips, do **not** start CR-88 against the flat paths and
plan to migrate later — the path mirroring between `messages/` and
`data/` is the whole point of §1a/§1b and is fragile to retrofit.

---

## 2. Facet groups (the canonical 19-column matrix)

Four facet groups, 19 values total, identical across all four sections.
These are the only dimensions the recommendation engine scores against.

| Group | Value `id`s (canonical lowercase keys) | Wizard chip labels | `CreateFlowState` field |
| --- | --- | --- | --- |
| `size` | `oneMember`, `twoToFive`, `sixToTwelve`, `thirteenToOneHundred`, `oneHundredToOneHundredK` | `1 member`, `2-5 members`, `6-12 members`, `13-100 members`, `100-100,000 members` | `selectedCommunitySizeIds` |
| `orgType` | `dao`, `forProfit`, `nonprofit`, `openSource`, `mutualAid`, `workersCoop` | `DAO`, `For profit business`, `Nonprofit`, `Open source project`, `Mutual aid`, `Worker's coop` | `selectedOrganizationTypeIds` |
| `scale` | `global`, `national`, `regional`, `local` | `Global`, `National`, `Regional`, `Local` | `selectedScaleIds` |
| `maturity` | `earlyStage`, `growthStage`, `established`, `enterprise` | `Early stage`, `Growth stage`, `Established`, `Enterprise` | `selectedMaturityIds` |

**Wizard chip ids vs facet value keys.** Wizard chips today use positional
1..N ids inside their messages files (see `chipRowsFromLabels` in
`app/(app)/create/screens/select/CommunityStructureSelectScreen.tsx` and
`CommunitySizeSelectScreen.tsx`). The recommendation layer needs a
**stable lookup** from those positional chip ids to the canonical facet
value keys above. Live that lookup table in
`data/create/customRule/_facetGroups.json` so every consumer sees the
same mapping.

Shape: a single object keyed by group, then by the canonical value id,
with the wizard chip's positional id and source messages path. The
chip's display label is **not** duplicated here — it stays in the
messages file (the lookup is just position → canonical id). Order in
each group's `values` follows §2's canonical key order, **not** the
chip's positional order in the messages file:

```json
{
  "size": {
    "source": "messages/en/create/community/communitySize.json#/communitySizes",
    "values": {
      "oneMember":               { "chipId": "1" },
      "twoToFive":               { "chipId": "2" },
      "sixToTwelve":             { "chipId": "3" },
      "thirteenToOneHundred":    { "chipId": "4" },
      "oneHundredToOneHundredK": { "chipId": "5" }
    }
  },
  "orgType": {
    "source": "messages/en/create/community/communityStructure.json#/organizationTypes",
    "values": {
      "workersCoop":  { "chipId": "1" },
      "mutualAid":    { "chipId": "2" },
      "openSource":   { "chipId": "3" },
      "nonprofit":    { "chipId": "4" },
      "forProfit":    { "chipId": "5" },
      "dao":          { "chipId": "6" }
    }
  },
  "scale": {
    "source": "messages/en/create/community/communityStructure.json#/scaleOptions",
    "values": {
      "local":    { "chipId": "1" },
      "regional": { "chipId": "2" },
      "national": { "chipId": "3" },
      "global":   { "chipId": "4" }
    }
  },
  "maturity": {
    "source": "messages/en/create/community/communityStructure.json#/maturityOptions",
    "values": {
      "earlyStage":  { "chipId": "1" },
      "growthStage": { "chipId": "2" },
      "established": { "chipId": "3" },
      "enterprise":  { "chipId": "4" }
    }
  }
}
```

Validated by the same Zod schema module as the facet files
(`lib/server/validation/methodFacetsSchemas.ts`). The parity test in §12
asserts every `chipId` matches an actual position in the referenced
messages file (off-by-one fails loudly). Adding a chip in messages
without updating this file → schema error.

---

## 3. Method inventory per section

Every method's slug must exist as a `methods[].id` in the corresponding
messages file *and* as a key in the corresponding facet file. Parity is
enforced at two points (no app-boot hook):

- **`prisma db seed`** — `loadSectionFacets()` runs the Zod schema
  against each `data/create/customRule/<section>.json` file before
  upserting; any orphan slug or unknown facet value fails the seed.
- **`tests/unit/methodFacets.test.ts`** — runs in CI on every PR;
  re-asserts the same parity statically (no DB needed) so authoring
  errors are caught before the seed ever runs.

There is intentionally **no Next.js app-boot validator** — schema
failures should surface in CI/seed, not at request time.

| Section | Screen | Messages file | Facet file | Method count |
| --- | --- | --- | --- | --- |
| `communication` | `app/(app)/create/screens/card/CommunicationMethodsScreen.tsx` | `messages/en/create/customRule/communication.json` | `data/create/customRule/communication.json` | 11 |
| `membership` | `app/(app)/create/screens/card/MembershipMethodsScreen.tsx` | `messages/en/create/customRule/membership.json` | `data/create/customRule/membership.json` | 19 |
| `decisionApproaches` | `app/(app)/create/screens/right-rail/DecisionApproachesScreen.tsx` | `messages/en/create/customRule/decisionApproaches.json` | `data/create/customRule/decisionApproaches.json` | 32 |
| `conflictManagement` | `app/(app)/create/screens/card/ConflictManagementScreen.tsx` | `messages/en/create/customRule/conflictManagement.json` | `data/create/customRule/conflictManagement.json` | 19 |

`section` keys above are the canonical lowercase camelCase tokens used in
the DB, the API query string, and the Zod schemas. The four xlsx
authoring artifacts in `~/Downloads/` (see "Source workbooks" appendix
below) are the human reference for which methods match which facets;
they do not ship.

---

## 4. Existing data model & wizard surface area

### 4.1 `RuleTemplate` (today, unchanged)

```64:73:prisma/schema.prisma
model RuleTemplate {
  id          String   @id @default(cuid())
  slug        String   @unique
  title       String
  category    String?
  description String?
  body        Json
  sortOrder   Int      @default(0)
  featured    Boolean  @default(false)
}
```

`body` JSON is the rendered rule document
(`{ sections: [{ categoryName, entries: [{ title, body }] }, ...] }`),
authored today by `bodyFromXlsxComposition()` in `prisma/seed.ts` from
the hand-typed `COMPOSITION_BY_SLUG` map. **Stays hand-curated** — the
matrix doesn't change how templates are authored, only how they are
ranked. Section ordering (Values → Communication → Membership →
Decision-making → Conflict management) is set by `governancePatternBody`
and is canonical.

### 4.2 Wizard facets captured today (`CreateFlowState`)

```ts
selectedCommunitySizeIds?: string[];
selectedOrganizationTypeIds?: string[];
selectedScaleIds?: string[];
selectedMaturityIds?: string[];
selectedCoreValueIds?: string[];
selectedCommunicationMethodIds?: string[];
selectedMembershipMethodIds?: string[];
selectedDecisionApproachIds?: string[];
selectedConflictManagementIds?: string[];
```

The first four are exactly the four facet groups in §2. The last four
are the user's chosen methods per section (used to pre-rank or to feed
recommendations on later steps). `createFlowStateSchema` in
`lib/server/validation/createFlowSchemas.ts` is the canonical Zod schema
— recommendation-side schemas should import the facet-id arrays from
there rather than redefine them.

### 4.3 Wizard step order

Source of truth: `app/(app)/create/utils/flowSteps.ts` (`FLOW_STEP_ORDER`).
Relevant slice for the matrix:

```
review → core-values → communication-methods → membership-methods →
decision-approaches → conflict-management → confirm-stakeholders → final-review
```

### 4.4 Where templates surface in the UI (in scope for ranking)

| Surface | File |
| --- | --- |
| Marketing home "Popular templates" | `app/(marketing)/_components/MarketingRuleStackSection.tsx` |
| Templates index | `app/(marketing)/templates/page.tsx` |
| Template preview (by slug) | `app/(app)/create/review-template/[slug]/page.tsx` |
| "Use without changes" → publish | `app/(app)/create/CreateFlowLayoutClient.tsx` `handleUseTemplateWithoutChanges` |
| API list | `app/api/templates/route.ts` (GET only, no params today) |

Template ranking adds optional facet query params to `/api/templates`;
the no-facets path keeps today's curated ordering. The
`/create/informational?template=<slug>` query-param prefill is a known
no-op (`CreateFlowLayoutClient.tsx`); fixing it is **out of scope**.

---

## 5. Repo conventions to follow (don't reinvent)

References point at the canonical example for each.

### 5.1 API routes (`app/api/**/route.ts`)

`app/api/drafts/me/route.ts` is the reference. Every new route in this
feature must match:

1. `if (!isDatabaseConfigured()) return dbUnavailable();` — always first
   (`lib/server/env.ts`, `lib/server/responses.ts`).
2. For auth'd routes: `const user = await getSessionUser();` then
   `NextResponse.json({ error: "Unauthorized" }, { status: 401 })` if
   missing. Recommendation read endpoints stay unauthenticated.
3. Request bodies: `readLimitedJson(request)` →
   `<schema>.safeParse(parsed.value)` → `jsonFromZodError(validated.error)`
   on failure (`lib/server/validation/requestBody.ts`,
   `lib/server/validation/zodHttp.ts`).
4. Success: `NextResponse.json({ <key>: data })` — flat object, no
   `success: true` envelope.
5. Errors: structured `{ error: { code, message } }` (Zod path) or simple
   `{ error: "..." }` (auth path).
6. Server-side query helpers swallow Prisma failures and return
   `[]`/`null` (see `listRuleTemplatesFromDb` in
   `lib/server/ruleTemplates.ts`). Routes do **not** wrap helper calls
   in `try/catch`.

### 5.2 Zod schemas live in `lib/server/validation/`

- One file per feature area
  (`lib/server/validation/methodFacetsSchemas.ts` for this feature).
- Export the schema **and** the inferred type
  (`export type X = z.infer<typeof xSchema>`).
- Reuse facet-id arrays from `createFlowStateSchema` rather than
  redeclaring them.

### 5.3 Prisma access

- Singleton: `import { prisma } from "lib/server/db";` — never
  `new PrismaClient()` from app code.
- Server-only fetch/list helpers live under `lib/server/<feature>.ts`,
  return DTOs (not raw Prisma rows), and degrade gracefully
  (`isDatabaseConfigured()` short-circuit, `try/catch` → empty result).
- Use `prisma.$transaction` for the seed step that swaps facet rows
  (delete-then-create per `(section, slug)` pair atomically).

### 5.4 DTO style

Hand-written `type` aliases that mirror a Prisma `select` clause,
co-located with the consumer (see `RuleTemplateDto` in
`lib/create/fetchTemplates.ts`).

### 5.5 Tests

- Vitest under `tests/unit/*.test.ts` for schemas and pure functions
  (see `tests/unit/createFlowValidation.test.ts`).
- Test the facet JSON files themselves with a parity test
  (`tests/unit/methodFacets.test.ts`) that loads each `data/` file
  alongside its messages file and asserts every `methods[].id` has a
  facet entry and vice-versa.
- API routes are not unit-tested today; cover behavior indirectly via
  schema tests.

### 5.6 Logging

Use `logger` from `lib/logger.ts` for any server-side info/warn/error in
scripts and route helpers. No `apiError` helper exists; do not introduce
one.

### 5.7 i18n stays the source of truth for copy

Card decks and modal copy live in
`messages/en/create/customRule/<section>.json` (post-reorg) and are read
via `useMessages().create.customRule.<section>` (`messages/en/index.ts`,
`app/contexts/MessagesContext.tsx`). The matrix never puts copy in the
DB. The recommendation API returns slugs and scores only — never copy —
so Storybook stories, MSW handlers, and Vitest specs keep importing the
message JSON statically and don't need an API mock for content.

### 5.8 Component-prop normalization rule

The repo's lowercase prop normalization rule
(`.cursor/rules/component-props.mdc`) applies to **React component props
only**. It does **not** apply to API query params, DB columns, JSON
keys, or messages keys. Conventions used by this feature: facet group
ids and facet values are camelCase (`orgType`, `workersCoop`); messages
keys are camelCase (`stepByStepInstructions`, `restorationFallbacks`);
method slugs are kebab-case (`peer-mediation`, `lazy-consensus`).

---

## 6. Authoring contract (the JSON shape)

Each `data/create/customRule/<section>.json` file is a single object
keyed by method slug. Each entry has the four facet groups, each holding
booleans (or numeric weights, optional) per canonical value id.

```ts
type FacetMatch = boolean | { match: boolean; weight?: number };

type SectionFacets = Record<
  string, // method slug, e.g. "peer-mediation"
  {
    size:     Record<"oneMember" | "twoToFive" | "sixToTwelve" | "thirteenToOneHundred" | "oneHundredToOneHundredK", FacetMatch>;
    orgType:  Record<"dao" | "forProfit" | "nonprofit" | "openSource" | "mutualAid" | "workersCoop", FacetMatch>;
    scale:    Record<"global" | "national" | "regional" | "local", FacetMatch>;
    maturity: Record<"earlyStage" | "growthStage" | "established" | "enterprise", FacetMatch>;
  }
>;
```

Example (illustrative, not a real entry):

```json
{
  "peer-mediation": {
    "size":     { "oneMember": false, "twoToFive": true,  "sixToTwelve": true,  "thirteenToOneHundred": true,  "oneHundredToOneHundredK": false },
    "orgType":  { "dao": false, "forProfit": false, "nonprofit": true, "openSource": true, "mutualAid": true, "workersCoop": true },
    "scale":    { "global": false, "national": false, "regional": true, "local": true },
    "maturity": { "earlyStage": true, "growthStage": true, "established": true, "enterprise": false }
  }
}
```

Bulk shorthand: a field that's omitted defaults to `false`, so files
only need to enumerate matching facets when authors prefer the compact
form. The Zod schema (run at seed time and in the parity test, per §3)
fills defaults and enforces:

- Every key in the file matches a `methods[].id` in the corresponding
  messages file (no orphans either way — fail loudly).
- Every facet group is present for every method (or fully omitted, in
  which case it defaults to "all false").
- Facet values are exactly the canonical ids in §2 (typos fail).

### One-time transcription from the workbooks

The four `.xlsx` files in `~/Downloads/` (see appendix) carry the facet
matches as the trailing 19 columns after the descriptive copy columns
(slug, label, supportText, applicableScope, consensusLevel, etc.). The
implementing agent should:

1. Inspect one workbook with `openpyxl` to confirm the column order
   matches §2 (`size` × 5, `orgType` × 6, `scale` × 4, `maturity` × 4 =
   19 facet columns). The copy ingest already confirmed the leading
   columns; the trailing 19 are the only new bit.
2. Write a one-shot Python script (mirror of the throwaway
   `/tmp/ingest_methods.py` used for the messages ingest — **do not
   commit**) that:
   - Reads each workbook's `Current` sheet,
   - Slugifies the first column the same way the messages ingest did
     (kebab-case, ASCII-folded, lowercase) so keys match
     `methods[].id` exactly,
   - Treats `✓` (and `1`, `true`, `yes` — case-insensitive) as `true` and
     everything else (`x`, blank, `-`, `0`, `false`, `no`, `✗`) as `false`.
     The workbooks intentionally use `✓` to mark a match and `x` to mark a
     non-match (cross-out) — both are filled cells but only `✓` counts as
     a positive recommendation,
   - Emits one `data/create/customRule/<section>.json` per workbook in
     the §6 shape, sorted by method slug,
   - Emits `data/create/customRule/_facetGroups.json` from the column
     header → canonical-id mapping above (§2).
3. Hand-review the diff against the workbook for the first ~3 methods
   per section, then commit the JSON files.
4. Run `tests/unit/methodFacets.test.ts` (§12) to confirm parity with
   the messages files, and `prisma db seed` to confirm the schema +
   transaction work.

The script is throwaway. Future facet edits happen by hand in the JSON
files. The workbooks stay at `~/Downloads/` and are **not** committed —
the post-ingest JSON files (both `messages/` and `data/`) are the
historical record.

---

## 7. Storage

Facet data lives in the JSON files above (the source of truth) and is
**hydrated into the DB at seed time** so the API can do efficient joins
when ranking templates by composed-method match. The JSON is canonical —
the DB is a derived index.

```prisma
model MethodFacet {
  id       String  @id @default(cuid())
  section  String  // communication | membership | decisionApproaches | conflictManagement
  slug     String  // matches the `id` of an entry in the messages file's `methods` array
  group    String  // size | orgType | scale | maturity
  value    String  // e.g. "workersCoop"
  matches  Boolean // ✓ → true, blank/explicit-false → false
  weight   Float?  // optional numeric override for future scoring tweaks
  @@unique([section, slug, group, value])
  @@index([section])
  @@index([group, value, matches])
}
```

Why this shape:

- **JSON is the source of truth** — diffs are reviewable in PRs, no
  binary artifacts, no importer to maintain. `prisma db seed` (or the
  facet-only equivalent script) reads the JSON and upserts rows.
- **DB enables joins** — template ranking joins `MethodFacet` against
  the methods listed in `RuleTemplate.body` to compute per-template
  match scores. In-memory ranking would also work but the join is
  cleaner and matches existing patterns.
- **Slug is the contract** — every JSON key matches a `methods[].id` in
  messages; boot-time validation enforces this both ways.
- **No runtime spreadsheet dependency** — `xlsx` / SheetJS is **not**
  added to `package.json`. The original `~/Downloads/*.xlsx` files
  authored both the messages content (already ingested) and the facet
  matches (one-time transcription into the JSON files described in §6).
  Future facet edits happen directly in the JSON.

---

## 8. Seed / sync

`prisma/seed.ts` (or a small co-located helper at
`prisma/seed/methodFacets.ts`) does:

```ts
async function seedMethodFacets() {
  const sections: Section[] = [
    "communication",
    "membership",
    "decisionApproaches",
    "conflictManagement",
  ];

  for (const section of sections) {
    const facets = await loadSectionFacets(section); // reads + Zod-validates the JSON
    await prisma.$transaction([
      prisma.methodFacet.deleteMany({ where: { section } }),
      prisma.methodFacet.createMany({
        data: flattenSectionFacets(section, facets),
      }),
    ]);
  }
}
```

- **Validation runs first.** `loadSectionFacets` reads the JSON and runs
  the Zod schema against it; any orphan slug, missing facet group, or
  unknown facet value fails the seed.
- **Per-section atomic swap.** Delete-then-create inside one
  transaction so the DB is never partially populated.
- **No flags, no `--dry-run`, no fixture workbooks.** The seed is
  idempotent and cheap (~80 methods × 19 facets = ~1500 rows total);
  re-running on every prisma seed is fine.
- **No app-boot validator.** Authoring errors surface in CI (parity
  test, §3) or at `prisma db seed` time — never at request time.

---

## 9. APIs

Both endpoints follow §5.1 conventions. **Neither returns copy** — copy
lives in messages and is read client-side via `useMessages()`.

### 9.1 `GET /api/templates` (rewrite)

Optional facet query params:

- `facet.size=<valueId>` (repeatable)
- `facet.orgType=<valueId>`
- `facet.scale=<valueId>`
- `facet.maturity=<valueId>`

Behavior:

- No params → existing curated ordering (`featured`, `sortOrder`,
  `title`), no scoring.
- With facets → score each template by joining its `body`-referenced
  method slugs against `MethodFacet` and **summing matches**, then
  sort highest-first.

**Scoring algorithm (simple count, v1).**

For each template, build the set of `(section, slug)` pairs from
`RuleTemplate.body` (the existing curated composition). For each
`(section, slug)` pair, count how many of the requested
`facet.<group>=<value>` query params match an existing
`MethodFacet { matches: true }` row. The template's score is the
**sum across all its methods**:

```
score(template)
  = Σ over (section, slug) in template.body:
      Σ over (group, value) in requested facets:
        1 if MethodFacet exists with matches=true for (section, slug, group, value)
        else 0
```

Notes:
- A template with five matching methods scores ~5× a template with one
  matching method, which is the desired bias toward whole-stack fit.
- `weight` in `MethodFacet` is **ignored** by v1 — it's only there so
  authors can store nuance for a future weighted-rank pass without a
  migration.
- Ties broken by today's curated ordering (`featured` desc → `sortOrder`
  asc → `title` asc) so no-facets and zero-match cases produce identical
  output to the existing endpoint.
- Templates with `score = 0` are still returned, ranked last by the
  curated tie-break (recommendations rank, never hide; see §10).

Response:

```ts
{
  templates: RuleTemplateDto[],
  scores?: Record<string, { score: number; matchedFacets: string[] }>
}
```

`scores[slug].matchedFacets` is the deduped list of
`"<section>:<slug>:<group>:<value>"` keys that contributed, useful for
debugging and for an eventual "Why this template?" UI tooltip.

### 9.2 `GET /api/create-flow/methods?section=<section>[&facet.*=...]`

Powers the four card-deck wizard steps. Returns slugs + per-method match
scores only — wizard renders by looking up entries in
`useMessages().create.customRule.<section>.methods` (via the
`methodById` map each screen builds).

Response:

```ts
{
  section: "communication" | "membership" | "decisionApproaches" | "conflictManagement",
  methods: Array<{
    slug: string;
    matches: { score: number; matchedFacets: string[] };
  }>
}
```

**Scoring algorithm.** Same simple count as §9.1, scoped to a single
method:

```
score(method)
  = Σ over (group, value) in requested facets:
      1 if MethodFacet exists with matches=true for (section, method.slug, group, value)
      else 0
```

Methods are returned **ranked by `matches.score` desc**, then by the
on-disk order from the messages file (so the deck stays stable when no
facets are passed and zero-match methods preserve authoring order). The
wizard never **hides** rows — see §10.

Server helper: `listMethodRecommendations({ section, facets })` in
`lib/server/methodRecommendations.ts`. Same swallow-and-return-`[]`
failure mode as `listRuleTemplatesFromDb`. When the DB is unavailable
(or facets are empty), the wizard falls back to the messages deck in
its on-disk order.

### 9.3 `POST /api/templates/recommend` (follow-up, optional)

If product wants to send the full `CreateFlowState` instead of just
facet ids, body schema reuses `createFlowStateSchema`. Skip until §9.1
+ §9.2 ship.

**Empty / partial facets:** never error. Fall back to today's ordering
and return all rows.

---

## 10. Wizard wiring (UI follow-on)

Once the API exists:

- `communication-methods` / `membership-methods` / `decision-approaches`
  / `conflict-management` screens each call
  `GET /api/create-flow/methods?section=...&facet.*=...` to get the
  ranking. Card label, description, and modal copy continue to come
  from `useMessages().create.customRule.<section>.methods` (a flat
  array — each screen already builds a `methodById` lookup map and
  iterates the array; no per-section `_CARD_ORDER` constants exist).
  The screen reorders the array by the API's ranked slug list before
  rendering.
- API failure or empty facets → render the messages deck in its on-disk
  order. No regression from today.
- Selecting a template on the marketing home or `templates/` page can
  prefill the create flow's `selected*MethodIds` from the template's
  composition (closes the `?template=` no-op gap noted in
  `CreateFlowLayoutClient.tsx`). Out of scope for CR-88.
- Recommendations **never hide** options — ranking only. Authors expect
  to see "all 32 decision-making patterns" with the matching ones
  surfaced first.

---

## 11. Resolved decisions (no open questions)

- ~~Where card copy lives~~ → `messages/en/create/customRule/*.json`,
  flat `methods` array per file. Done.
- ~~Card / modal split in messages files~~ → collapsed into a single
  `methods` array; modal title/description derive from each entry's
  `label`/`supportText`. Done.
- ~~`rightRail.json` rename~~ → file is now
  `messages/en/create/decisionApproaches.json`; namespace is
  `m.create.decisionApproaches` (and will become
  `m.create.customRule.decisionApproaches` after the §1c folder reorg).
  Done.
- ~~Facet authoring format~~ → typed JSON files committed under
  `data/create/customRule/`, validated by Zod. No spreadsheets at
  runtime, no `xlsx` dep, no importer.
- ~~Where facet data lives at runtime~~ → JSON is canonical; DB is
  hydrated at seed time for join-friendly queries.
- ~~Decision-making `Consensus Level` scale~~ → integer 0-100 in
  messages; the original spreadsheet's 0.0-1.0 floats were converted
  during the one-time content ingest.
- ~~Membership section key naming~~ → `eligibility` /
  `joiningProcess` / `expectations` (matches wizard).
- ~~Scoring vs filtering~~ → ranking only; never hide rows (§10).
- ~~`RuleTemplate` rows~~ → stay hand-curated in `prisma/seed.ts`
  `COMPOSITION_BY_SLUG`. The matrix just adds ranking; it doesn't
  regenerate templates.
- ~~Values matrix~~ → out of scope. Values are baked into each
  curated template (and authored in `coreValues.json` for the
  open-ended wizard step). No facet matrix needed; if a template is
  recommended, its values come along statically.
- ~~Ranking algorithm~~ → simple count (sum of `MethodFacet { matches:
  true }` rows touched by the requested facets); per-method for §9.2
  and per-template (sum across composed methods) for §9.1. `weight` is
  reserved for a future v2; ignored by v1.
- ~~Boot-time validation~~ → none. Parity is enforced by the seed step
  (§8) and the parity test in CI (§3, §12). No `next dev` startup hook.
- ~~Messages folder reorg sequencing~~ → ships as **its own ticket
  before** CR-88 (§1c). CR-88 assumes the post-reorg paths.
- ~~Spreadsheet handoff~~ → the four `~/Downloads/*.xlsx` files are
  passed to the implementing agent alongside this doc. They are **not**
  committed; the post-ingest `messages/en/create/customRule/*.json`
  and `data/create/customRule/*.json` files are the historical record.

---

## 12. Test plan (acceptance for CR-88)

- [ ] `prisma db seed` populates `MethodFacet` from the four
      `data/create/customRule/<section>.json` files with no errors,
      producing the expected row count
      (`(11 + 19 + 32 + 19) × 19 = 1539` rows max, fewer if authors
      use the omit-default shorthand).
- [ ] `tests/unit/methodFacets.test.ts` asserts every method slug in
      each facet file matches a `methods[].id` in the corresponding
      messages file (and vice-versa) — no orphans either way. Also
      asserts every `chipId` in `_facetGroups.json` resolves to a real
      position in the referenced messages file (off-by-one fails).
- [ ] `tests/unit/methodFacetsSchemas.test.ts` exercises the Zod schema
      (rejects unknown facet values, unknown groups, unknown sections,
      malformed booleans).
- [ ] `tests/unit/methodRecommendations.test.ts` exercises the scoring
      function directly with a fixture set: a method matching 2 of 3
      requested facets scores `2`; a template composing two methods
      that each match `2` and `3` requested facets scores `5`; ties
      fall back to curated `(featured, sortOrder, title)` order.
- [ ] `GET /api/create-flow/methods?section=conflictManagement&facet.orgType=nonprofit`
      returns all 19 methods, ranked, with the `nonprofit`-matching
      methods scoring higher than non-matching ones; zero-match
      methods preserve their on-disk authoring order.
- [ ] `GET /api/templates?facet.orgType=nonprofit&facet.size=sixToTwelve`
      returns templates re-ordered by composed-method match count, with
      score-0 templates still present at the end in curated order.
- [ ] No-facets `GET /api/templates` matches today's curated ordering
      (no regression for the existing marketing/templates surfaces).
- [ ] DB-down smoke: with `DATABASE_URL` unset, the four wizard
      card-deck steps still render the full deck from messages (no
      5xx, no broken cards).
- [ ] Editing a `data/create/customRule/<section>.json` entry and
      re-running `prisma db seed` changes the rank order returned by
      both endpoints without any code change.

---

## 13. Source files referenced

- `prisma/schema.prisma` — `RuleTemplate` model (unchanged); add
  `MethodFacet` model (§7).
- `prisma/seed.ts` — current curated composition; add `seedMethodFacets`
  helper (§8).
- `app/api/templates/route.ts` — existing GET endpoint (rewrite with
  optional facet params).
- `app/api/drafts/me/route.ts` — reference route shape.
- `lib/server/db.ts` — Prisma singleton.
- `lib/server/responses.ts` — `dbUnavailable()`.
- `lib/server/ruleTemplates.ts` — `listRuleTemplatesFromDb` (extend with
  facet param + scoring helper).
- `lib/server/methodRecommendations.ts` — **new**; helper for §9.2.
- `lib/server/validation/methodFacetsSchemas.ts` — **new**; Zod schema
  for the JSON facet files and the API request shapes.
- `lib/server/validation/createFlowSchemas.ts` — reuse facet-id arrays
  rather than redeclaring them.
- `lib/server/validation/requestBody.ts` — `readLimitedJson`.
- `lib/server/validation/zodHttp.ts` — `jsonFromZodError`.
- `lib/logger.ts` — server-side `logger`.
- `app/(app)/create/types.ts` — `CreateFlowState` and facet fields.
- `app/(app)/create/utils/flowSteps.ts` — canonical step order.
- `app/(app)/create/utils/createFlowScreenRegistry.ts` — screen
  metadata.
- `app/(app)/create/screens/select/CommunityStructureSelectScreen.tsx`
  — chip-id derivation pattern (positional `String(i+1)`).
- `app/(app)/create/screens/card/CommunicationMethodsScreen.tsx` (and
  the three sibling screens) — already iterate `methods[]` via
  `methodById`; the API ranking layer plugs in here.
- `messages/en/create/customRule/{communication,membership,decisionApproaches,conflictManagement}.json`
  — flat `methods` arrays (post-reorg paths). Source of truth for copy;
  the matrix never edits these.
- `messages/en/create/{community,reviewAndComplete}/*.json` — the other
  two stages (post-reorg); not consumed by the matrix but listed for
  context on the §1c reorg.
- `data/create/customRule/{communication,membership,decisionApproaches,conflictManagement}.json`
  — **new**; facet matches per method.
- `data/create/customRule/_facetGroups.json` — **new**; canonical facet
  group/value ids and the wizard-chip-id ↔ facet-value-id mapping.
- `tests/unit/createFlowValidation.test.ts` — Vitest pattern for new
  schema/parity tests.
- Roadmap: `docs/guides/backend-roadmap.md` §4, §13.
- Spec: `docs/guides/backend-linear-tickets.md` Ticket 16.

---

## Appendix — Source workbooks (one-time authoring artifact)

These four spreadsheets are **handed to the implementing agent
alongside this doc**. They were used once to seed the messages content
(already done) and will be used once more to transcribe the facet
matches into `data/create/customRule/*.json` per §6's "One-time
transcription" steps. They are **not** committed to the repo, **not**
part of the runtime contract, and **not** referenced by any code path.

| File (in `~/Downloads/`) | Sheet | Rows | Section |
| --- | --- | --- | --- |
| `Communication Methods.xlsx` | `Current` | 11 | `communication` |
| `Group_Membership_Methods.xlsx` | `Current` | 19 | `membership` |
| `Decision-making.xlsx` | `Current` | 32 | `decisionApproaches` |
| `Conflict Management Methods.xlsx` | `Current` | 19 | `conflictManagement` |

Each workbook's leading columns hold the descriptive copy already
ingested into `messages/en/create/customRule/<section>.json`; the
trailing 19 columns hold the facet matches that need to land in
`data/create/customRule/<section>.json`. After CR-88 lands, future
facet edits happen directly in the JSON files — the workbooks are
historical reference only, and the committed JSON (in both `messages/`
and `data/`) is the canonical record.