UniKit Rules — Project Knowledge Base Router

Manage project rules in .unikit/memory/. This skill is a module-agnostic router: the knowledge base is partitioned into top-level modules (laid out on disk as .unikit/memory/<module>/<tier>), and this skill resolves which module a request targets, loads that module's contract, and then drives the generic add/research/migrate/validate workflows against the module's tiers.

The list of registered modules and their structural fields (tiers, enginePartitioned, skillPrefix) comes from .unikit/system/modules.yml. Each module's content semantics, file format, and path conventions come from its contract file {{skills_dir}}/{{self_name}}/references/module-<id>.md.

Two modules are registered today: code (tiers core + stack) and gamedesign (tiers core + library) — read the live list from modules.yml, never from this paragraph. The router is written so additional modules can be registered later with no change to this file — only a new modules.yml entry and a new module-<id>.md contract.

What Belongs Here

• Content that belongs to one of the registered modules' tiers, as defined by that module's contract (references/module-<id>.md → "What Belongs Here" / "Content Classification").

What Does NOT Belong Here

• Architecture decisions (module boundaries, dependency flow, DI structure) → .unikit/ARCHITECTURE.md
• Project-specific overrides (explicit deviations from base rules) → .unikit/RULES.md via the unikit-rules skill

If the user provides content that falls into these categories, inform them and suggest the correct destination. For RULES.md content, suggest using the unikit-rules skill instead. The active module's contract may define additional "does not belong here" cases — honor them too.

Language Awareness — BLOCKING PRE-REQUISITE

BEFORE producing ANY output, silently read .unikit/system/LANGUAGE_RULES.md and apply its rules to ALL subsequent output. If the file is missing or unreadable, fall back to English. Do not produce any user-facing output until language rules are loaded. Do not announce, confirm, or mention the language setting.

The language holds for the whole session, not just at load time: every message until the conversation ends is in language.ui — progress notes while agents run, relays of what a subagent returned, the final report, any follow-up discussion. English input (subagent results, tool output, these instructions) is data, never a cue to switch languages.

Skill calls

Where this skill says to invoke <skill> — optionally "with the argument <text>" — it means one call, made in this session, now, on every runtime: not a command printed for the user. A call site that names no argument passes none.

1. If you have a Skill tool that accepts arguments, call it with that skill and the arguments of the call site, in full and unchanged, then follow the skill.
2. Otherwise — no Skill tool, or one that takes only a name — Read {{skills_dir}}/<skill>/SKILL.md in full, treat it as the instructions of this step and carry them out here, in this session, now, with the arguments of the call site as the skill's arguments.
3. When the skill has run, go on to the step the call site names. Do not print a command for the user to run and do not stop.
4. Only if that file cannot be read: print Run: /unikit-<skill> <arguments> for the user and stop.

Workflow

Step 0: Load Configuration (silent — do NOT print or announce any of these values)

1. Read .unikit/config.yaml — extract language.ui for user-facing messages (default: en)
2. Read .unikit/system/modules.yml — the authoritative module registry snapshot. Parse every entry under modules: into { id, tiers, enginePartitioned, skillPrefix }. This is the source of truth for which modules exist and what tiers each partitions its rules into. If the file is missing (older install), fall back to a single module code with tiers core, stack and surface a WARN suggesting unikit-ai update to regenerate it.
3. Read .unikit/skill-context/{{self_name}}/SKILL.md — MANDATORY if the file exists.

This file contains project-specific workflow rules added by /unikit-skills-context or /unikit-evolve. These rules change how this skill operates (research priorities, cross-check behavior, file format conventions, etc.).

How to apply skill-context rules:

• Treat them as project-level overrides for this skill's general instructions
• When a skill-context rule conflicts with a general rule written in this SKILL.md, the skill-context rule wins (more specific context takes priority)
• When there is no conflict, apply both: general rules from SKILL.md + project rules from skill-context
• Do NOT ignore skill-context rules even if they seem to contradict this skill's defaults — they exist because the project's experience proved the default insufficient

Do NOT output status messages like "Язык: X" or "Module: X". Proceed directly to Step 0.5.

Step 0.5: Resolve the Active Module

Determine which module this request targets before classifying input. Never guess silently when the target is ambiguous.

Resolution policy (in order):

1. Explicit --module <id> flag. Scan $ARGUMENTS for --module <id>. If present:
   • Validate <id> against the ids parsed from modules.yml.
   • Valid → set moduleId = <id>, strip the --module <id> tokens from $ARGUMENTS, and skip the remaining resolution steps.
   • Invalid (no such module) → report the unknown id and the list of registered module ids via AskUserQuestion (options = registered modules), then use the chosen id. Strip the flag tokens regardless. Print the module ids with their one-line domain summaries to the screen as plain markdown first — the question mechanism carries the options and nothing else.
2. No flag → infer from prompt context. Match the request (rule content, URL, file path, stated intent) against each registered module's domain. The domain of a module is described in its contract file {{skills_dir}}/{{self_name}}/references/module-<id>.md (its "What Belongs Here" / "Content Classification" sections) plus the modules.yml metadata. Read the candidate contract files as needed to judge fit.
3. Confident single match → use it silently. If exactly one module's domain plausibly covers the request, set moduleId to it and continue without asking. A neutral or code-oriented request when code is the only registered module resolves silently to code.
4. Ambiguity → ALWAYS ask. If zero modules fit, the content fits none of the registered modules, or two or more modules are plausible, ask the user via AskUserQuestion. Options = every registered module id (with a one-line domain summary) + a final "help me choose — I'll describe what this is" option. Never fall back to a default module to avoid the question. Print the module ids with their one-line domain summaries to the screen as plain markdown first — the question mechanism carries the options and nothing else.
   • Edge case with a single registered module: a request that is explicitly not about that module's domain (and fits no registered module) must still trigger the question — do not dump unrelated content into code just because it is the only module.

After resolution, the following are fixed for the rest of the run:

• moduleId — the resolved module id.
• tiers — that module's ordered tier list from modules.yml (for code: core, stack).
• Module contract — read {{skills_dir}}/{{self_name}}/references/module-<moduleId>.md now. It defines tier semantics, content classification, scope rules, and all file formats. Apply it throughout the rest of the workflow.
• Memory root — .unikit/memory/<moduleId>/.
• Tier directories — .unikit/memory/<moduleId>/<tier>/ for each tier.
• Index file — .unikit/memory/<moduleId>/RULES_INDEX.md. Read it now for the current index of all rule files in this module.

Everywhere below, <module> means moduleId and <tier> means one of the resolved module's tiers. Concrete tier names, the "what goes in which tier" decision, and the file-format templates are always taken from the module contract — this router never hardcodes them.

Step 1: Parse Input & Classify Intent

Two-phase classification determines both input type and user intent.

Phase A — Input type:

First, strip the already-parsed flags from $ARGUMENTS before classifying input type:
├── --skip-registry present → Remember skipRegistry=true, strip the flag (absent → false).
│   The remaining text is classified normally.
└── (The --module flag, if any, was already stripped in Step 0.5.)

Then classify the (possibly stripped) $ARGUMENTS:
├── First token is "optimise" or "optimize" (bare keyword, case-insensitive) → OPTIMISE
│   (jump to Branch E). Consume the keyword; any remaining text is the optional rule scope
│   (specific rule names/ids). An empty remainder means "ask which rules" — Branch E handles
│   it. OPTIMISE rewrites no content: it only relocates existing sections into reference files.
│   NOTE: "optimise" is a bare positional keyword matched on the FIRST token (not exact
│   equality, because it carries an optional scope) — there is NO "--optimise" flag form. A
│   description that genuinely begins with the word "optimise" is therefore read as this
│   intent; ask the user to rephrase if that was not the intent.
├── Equals "validate" (case-insensitive) → VALIDATE INDEX (jump to Branch D)
├── First token is "migrate-rules" (bare keyword) or "--migrate-rules" (legacy alias) →
│   ├── $ARGUMENTS is exactly "migrate-rules" or "--migrate-rules" (no other text) → MIGRATE RULES (jump to Branch C)
│   └── $ARGUMENTS has additional text besides the keyword →
│       Report to user: "migrate-rules must be used alone without additional arguments."
│       STOP. Do not proceed.
├── Matches RULES.md path (case-insensitive):
│   "RULES.md", "rules.md", ".unikit/RULES.md",
│   ".unikit/rules.md", or any path ending with /RULES.md
│   → MIGRATE RULES (jump to Branch C)
├── Contains BOTH http(s):// tokens AND file extension / path tokens → Mixed input (URL + File)
├── Starts with http:// or https:// → URL input (can be multiple, space-separated)
├── Ends with a known file extension (.md, .txt, .json, .yaml, .pdf, .fb2, .epub), or
│   names an existing file OR folder → File input (a folder, a PDF, an EPUB/FB2 book, or
│   an oversized text file is a LARGE source — see the note below; B.1 routes it through
│   large-sources.md, not a naive Read)
├── Has text content → Description input
└── No arguments → Interactive mode: ask user what to document, then re-classify

Large sources (folder / PDF / EPUB+FB2 book / oversized file). When the File-input branch matches a folder, a PDF, an EPUB or FB2 book (.epub, .fb2, .fb2.zip), or a text file too large to read in one pass, the RESEARCH pipeline does not Read it naively. The actual gathering is delegated to the large-source workflow at {{skills_dir}}/{{self_name}}/references/large-sources.md (TOC-first → topic-map → chunk → cleanup), which uses the probe-gated Python 3 helper {{skills_dir}}/{{self_name}}/scripts/material-prep.py when a Python 3 interpreter is available and falls back to a manual extraction strategy otherwise. Book formats are extracted with the standard library (no new dependencies); MOBI/AZW is recognized but rejected — ask the user for an EPUB export. Step 1 only classifies the input here; B.1 ("File input") performs the delegation.

The --skip-registry flag is an escape hatch for callers that have already performed a registry lookup at a higher level and do not want this skill to repeat it. When the flag is present, Step 1.5 is bypassed entirely and the skill proceeds directly to content classification (Step 2) and generation.

Phase B — Determine intent based on input type:

Intent?
├── OPTIMISE        → Branch E (always — retroactive reference extraction, no new content)
├── MIGRATE RULES   → Branch C (always — explicit migration request)
├── Mixed (URL + File) → RESEARCH (always — both inputs processed together)
├── URL(s)           → RESEARCH (always — URLs signal delegation)
├── File/Folder path → RESEARCH (always — user provides source material for analysis)
├── Description      → Analyze tone ↓
│   ├── Exploratory  → RESEARCH
│   └── Concrete     → ADD RULE
└── Interactive      → Ask user, then re-classify

Tone signals for Description input:

Signal	Intent	Examples
Vague topic, delegation words	Research	"add rules for DOTween", "document how we use Addressables", "figure out best practices for R3"
Imperative, specific convention, code examples	Add Rule	"add rule: always use .SetEase(Ease.OutQuad)", "never use coroutines", "use NonLazy() for controllers"

Phase C — Detect integration intent (contract-defined):

If the active module's contract defines integration-gated tiers, detect integrationIntent per the contract's "Integration Intent" section and store it alongside the classified intent to thread into Branch A / Branch B. The code contract defines this for its stack tier (detection signals T1/T2, plus T3 post-fetch in the Research pipeline). Modules with no integration-gated tier skip this phase and leave integrationIntent=false.

Step 1.5: Registry-first Lookup

Always try the registry before generating a rule from scratch. If a matching rule already exists in the remote catalog, the user gets a vetted version by installing it via CLI — much better than spending tokens generating a local copy that duplicates upstream work.

1.5.1: Read CLI contract

Read .unikit/system/cli-contract.md to confirm available unikit-ai rules * commands and exit codes.

1.5.2: Fetch catalog and installed state

Run two CLI commands (sequential is fine). Always pass --module <moduleId> — the router already resolved the active module (Step 0.5), and every machine consumer sends its --module so the CLI returns the stable flat-single form no matter how many modules the registry carries (no---module rules list now defaults to all modules blocked + a flat-all JSON, which this skill does not want):

unikit-ai rules list --module <moduleId> --json

unikit-ai rules status --module <moduleId> --json

rules list --module <moduleId> --json returns the flat-single shape { engine, module, rules: [...] } — parse catalog.rules exactly as before (the scoped form carries no per-row module key, so the parser is unchanged). rules status does accept --module (it scopes the reported rows to that module); pass it so status.rules[] is already limited to the active module — no client-side filtering needed.

Exit code handling:

• rules list exit 2 (registry chain unreachable) → skip to Step 2 silently, proceed without registry
• rules list exit 1 → strictly "no .unikit.json" (not a UniKit project) → skip to Step 2. Engine-missing is no longer exit 1: a valid project whose engine is absent from the registry now exits 0 with an empty catalog.rules, which flows normally into Step 2.
• rules list exit 3 (unknown --module) → the resolved module id is not registered; report the error and skip to Step 2
• any other non-zero → report the error and skip to Step 2

Parse rules list --module <moduleId> --json into catalog.rules[] and rules status --module <moduleId> --json into status.rules[].

1.5.3: Semantic matching

Match the user's intent against catalog.rules[]:

• Exact id match (case-insensitive): "DOTween" → dotween
• Description match: "tweening library" → dotween via the description field
• Synonyms and localized names: "контейнер внедрения зависимостей" → zenject

Collect 0..N candidates. If N > 1, treat all as possible matches for the user prompt below.

1.5.4: Join with installed state

For each candidate, look it up in status.rules[] by name (case-insensitive):

• Already installed → note source, version, and tier so the prompt can show accurate state.
• Not installed → candidate is offerable as a fresh install.

1.5.5: Decision tree

Match result?
├── Exactly 1 match + already installed
│   → Inform user: "{name} is already installed (source={source}, v{version})"
│   → AskUserQuestion: What should we do?
│   │  Options:
│   │  1. Leave as is         → STOP
│   │  2. Edit manually       → Continue to Step 2
│   │  3. Reinstall from registry → Bash: unikit-ai rules install <id> --force → STOP
│
├── Exactly 1 match + NOT installed
│   → AskUserQuestion: Matching rule "{name}" v{version} found in registry. Install it?
│   │  Options:
│   │  1. Yes, install from registry  → install path (below)
│   │  2. Preview before deciding     → preview path (below)
│   │  3. No, generate a new rule     → Continue to Step 2 (research/manual generation)
│   │
│   ├── install path:
│   │   Bash: unikit-ai rules install <id>
│   │   → Report "✓ Installed <id> from registry" and STOP. The CLI already
│   │     regenerates RULES_INDEX.md — no extra sync needed.
│   │
│   └── preview path:
│       Bash: unikit-ai rules show <id> --module <moduleId> --references
│       (rules show is module-agnostic: a bare id that collides across modules
│        exits 3, so the already-resolved router always scopes with --module)
│       → Show content to user, then AskUserQuestion again:
│         1. Install → install path above
│         2. Generate a new rule → Step 2
│
├── Multiple matches
│   → AskUserQuestion: present all candidates + a "none of these, generate a new rule" option
│   → If user picks a candidate → treat as "exactly 1 match" (installed or not)
│   → If "none" → Continue to Step 2
│
└── Zero matches
    → Continue to Step 2 (research/manual generation)

Key behavior guarantees:

• The Multiple matches branch prints the candidates (name, version, module) to the screen as plain markdown first — the question mechanism carries the options and nothing else.
• Agreeing to install always ends with rules install <id> and terminates the skill. The content in the registry is the source of truth; the skill does not second-guess it.
• Declining the registry version always routes to Step 2, where the skill generates the rule from the user's description, URL, or file input as usual. No partial install, no mix of registry + local content.

When to skip Step 1.5 entirely

Bypass the lookup completely when any of the following is true:

• --skip-registry flag was present in $ARGUMENTS (a higher-level caller already queried the catalog and does not want redundant prompts)
• Intent is VALIDATE INDEX (Branch D), MIGRATE RULES (Branch C), or OPTIMISE (Branch E) — none of them installs a new rule, so registry lookup is irrelevant
• Input is a file path — the user explicitly wants to process a specific file, not search the catalog
• cli-contract.md doesn't exist — the UniKit CLI is not available in this project, so rules * commands cannot be used

---

Step 2: Classify Content Category

Use the active module's contract (references/module-<moduleId>.md → "Content Classification") to:

1. Pick the target tier for the content among the module's tiers, or determine the content belongs to none of the module's tiers.
   • If the content fits none of the tiers, redirect the user to the correct destination named in the contract's "What Does NOT Belong" section (architecture → .unikit/ARCHITECTURE.md; project-specific overrides → .unikit/RULES.md via unikit-rules). Stop here. Do NOT create any files.
   • If the content fits none of the tiers and seems to belong to a different module than the one resolved in Step 0.5, return to Step 0.5 and re-ask the module question rather than forcing a tier.
2. Identify the target file within the chosen tier, following the contract's filename conventions.
3. Check whether the file already exists under .unikit/memory/<module>/<tier>/.

Step 3: Route by Intent

Based on intent (Step 1) and classification (Step 2), follow the appropriate branch:

Intent?
├── VALIDATE INDEX → Branch D (inline below; skip Steps 2-3 entirely)
├── OPTIMISE       → Branch E (inline below; skip Step 2 — it operates on existing rules)
├── MIGRATE RULES  → read references/migrate-rules.md and follow it (skip Step 2)
├── ADD RULE       → Branch A (inline below)
└── RESEARCH       → read references/research-pipeline.md and follow it

Branches B (Research) and C (Migrate) are intent-gated and live in dedicated reference files loaded relative to this skill ({{skills_dir}}/{{self_name}}/references/): open the matching file only when its intent is selected, then return here for the Final Step: Confirm. Branch A, Branch D, and Branch E stay inline below since they are short and on the common path.

---

Branch A: Add Rule

The user has a concrete rule to save. No research, no enrichment — just store what the user explicitly stated.

A.1: Check Target File

File exists?
├── YES → A.2 (Cross-Check & Append)
└── NO
    ├── Tier is framework-agnostic (per the module contract, e.g. `code`/core) → A.3 (Create File)
    └── Tier may need research (per the module contract, e.g. `code`/stack)     → A.4 (Ask About Research)

A.2: Cross-Check, Gap List & Append (file exists)

Before appending, compare the new content against existing rules:

1. Read the target file.
2. Read related files — scan RULES_INDEX.md for files with overlapping keywords, then read those too (across all tiers of the module). Also read .unikit/RULES.md for project-specific overrides. Rule topics: plus the topic files listed under its ## Topics whose Load when matches the new rule's area; when unsure, load.
3. Compare each new rule against existing formulations. A contradiction is when the new content prescribes something different from an existing rule — e.g., "use await directly" vs existing "always wrap in UniTask.Create", or "bind as Transient" vs existing "bind as Singleton".
4. Build the gap list (the default for every update). Classify the new material against what the file already says, and show the result to the user as the summary of what the update will do:
   • Add — genuinely new rules not yet present in the file.
   • Change — rules that refine or supersede an existing formulation (these are the contradictions resolved below).
   • Unchanged — material the file already covers; skip it, do not restate. This gap-list discipline is the canonical cross-check/update behaviour that research-pipeline.md B.4 defers to ("Same logic as A.2"). It runs on every existing-file update regardless of input.

If contradictions found:

• Report each contradiction to the user, quoting both the new statement and the existing rule with its source file.

• For each contradiction, ask via AskUserQuestion (batch up to 3 contradictions per question):

  Options:

  1. Replace — update existing rule with new formulation
  2. Keep existing — discard the conflicting part
  3. Keep both — document both with a note on when each applies

  Based on choice:

  • Replace → overwrite the existing rule text in the target file with the new formulation
  • Keep existing → discard the conflicting part from the new content, keep existing rule unchanged
  • Keep both → add the new rule alongside the existing one with a note clarifying when each applies

After resolution (or if no contradictions):

• Apply the gap list: append the Add items and apply the Change items, merging into the appropriate sections. Skip Unchanged material. Do not overwrite useful existing content.
• → Go to Final Step: Confirm

A.3: Create New File (framework-agnostic tier, no existing file)

1. Create the rule file following the File Format Template from the module contract.
2. Write the user's rule(s) as-is — no enrichment, no synthesis.
3. Update RULES_INDEX.md — add a new row to the appropriate table in alphabetical order.
4. → Go to Final Step: Confirm

A.4: Ask About Research (research-eligible tier, no existing file)

Ask the user: "I don't have a rules file for {framework/topic} yet. Would you like me to research this and create a comprehensive document, or just save your rule as-is?"

User answer?
├── YES (research) → read references/research-pipeline.md and follow it (start at B.1)
└── NO (save as-is)
    ├── Create the rule file with user's rule(s) only
    ├── Update RULES_INDEX.md — add to the tier's table
    └── → Go to Final Step: Confirm

---

Branch B: Research

Full research pipeline — gather material, enrich, synthesize, then write into the active module's tier. The complete workflow (B.1–B.5: gather → integration upgrade → Context7 → synthesize → reference candidates → cross-check → write) lives in {{skills_dir}}/{{self_name}}/references/research-pipeline.md. Read that file and follow it, then return here for the Final Step: Confirm. All file-format and scope decisions in it defer to the module contract.

---

Branch C: Migrate Rules from RULES.md

Transfer mature entries from .unikit/RULES.md (the quick-capture staging area owned by unikit-rules) into permanent rule files under .unikit/memory/<module>/. The complete workflow (C.1–C.4: load → classify → present candidates → rephrase → apply) lives in {{skills_dir}}/{{self_name}}/references/migrate-rules.md. Read that file and follow it, then return here for the Final Step: Confirm.

---

Branch D: Validate Index

Sync .unikit/memory/<module>/RULES_INDEX.md with the actual files on disk. No user interaction — fully automatic.

D.1: Scan Actual Files

1. For each <tier> of the resolved module: Glob: .unikit/memory/<module>/<tier>/*.md — collect all real rule files in that tier.
2. Read .unikit/memory/<module>/RULES_INDEX.md — parse every table (one per tier), extract filenames from each row.

D.2: Diff

Compare the two sets:

• Missing from index — files that exist on disk but have no row in RULES_INDEX.md
• Phantom in index — rows in RULES_INDEX.md whose files don't exist on disk

D.3: Fix Index

For each missing file — read the file, extract the > **Scope**: and > **Load when**: lines from the header. Add a row to the appropriate tier's table in RULES_INDEX.md (alphabetical order), following the contract's "RULES_INDEX.md Format":

| {FILENAME}.md | {Scope value} | {Load when value} |

If the file has no Scope/Load when header, use the first heading and first paragraph to generate a brief description and keywords.

For each phantom entry — remove the row from RULES_INDEX.md using Edit.

D.4: Report

Output a short validation report (in user's configured language):

RULES_INDEX.md validation complete.

{per-tier counts, e.g. Core: {N} files, Stack: {M} files}
✅ Added to index: {X}
🔄 Removed from index: {Y}

{If X > 0: list added filenames}
{If Y > 0: list removed filenames}

Stop. Do not proceed to any other step.

---

Branch E: Optimise (retroactive reference extraction)

Retroactively shrink existing rules by moving their large/optional sections into reference files — the same outcome the on-add reference step (B.3.5) produces, applied after the fact. Branch E synthesizes no new content: it relocates text that already lives in a rule, so a rule that grew too long (e.g. a distilled book that landed entirely inline) becomes a lean conceptual rule plus on-demand references. The active <module> and its <tier>s were already resolved in Step 0.5; the module contract's Candidate Analyzer is the engine.

This branch is non-destructive: content moves, it is never deleted, every proposal is confirmed before writing, and it leans on git as the safety net. It does not touch the registry, src/, or the helper script.

E.1: Determine scope

Parse the rule scope from the remaining $ARGUMENTS (after the optimise keyword was consumed in Phase A):

• Named rules present (e.g. optimise dotween zenject) → that is the scope. Resolve each name to a file under .unikit/memory/<module>/<tier>/ (match against RULES_INDEX.md; ask if a name is ambiguous).

• No names → AskUserQuestion:

  Options:

  1. Specific rules — I'll name them → ask which, then resolve as above
  2. All rules in this module → scope = every rule file across the module's tiers

Whatever the scope, process the rules in batches of 2-3 at a time — analyze, present, confirm, and write one batch before moving to the next. This keeps each confirmation focused and the diff reviewable.

E.2: Analyze each rule

For each rule in the current batch:

1. Read the rule file.
2. Run the module contract's Candidate Analyzer ("Reference Candidate Extraction") over its sections: score each block on size (≳ 40 lines / ≳ 1500 chars), optionality, and lookup shape, and sort candidates into Tier 1 (clear wins) and Tier 2 (borderline).
3. Run the merge guard (research-pipeline.md B.3.5): before proposing any new reference file, scan the tier's existing references/ for a file that already covers the topic and target it instead of forking a near-duplicate.

E.3: Present and confirm

Present the candidates per rule, grouped into Tier 1 and Tier 2 (each with which signals fired and the target .unikit/memory/<module>/<tier>/references/{file}.md), then AskUserQuestion exactly as B.3.5 does:

Options:

1. Extract Tier 1 only
2. Extract Tier 1 + Tier 2
3. Adjust — rename / merge / split / drop, or move a block between tiers → revise and ask again
4. None — leave this rule unchanged

Always report the result, including "0 candidates" for a rule that is already lean — Branch E must say so rather than silently skip it.

E.4: Move the content (never delete)

For each approved candidate:

1. Write the reference file following the contract's Reference File Format (create new, or update the existing matching file per the merge guard).
2. Cut the moved section from the main rule — the content is relocated, not duplicated. Leave the conceptual rules intact.
3. Add the file to the main rule's > **References**: header line and add/extend a ## {Content} Lookup Workflow section telling the LLM when to open it. The main rule keeps only pointers, not the moved data.

Reference files are not indexed in RULES_INDEX.md (only main rules are); the main rule's row is unchanged unless its Scope/Load when changed (they should not).

E.5: Finish

After the last batch, go to the Final Step: Confirm — it runs unikit-ai rules sync (reconciling the moved files into .unikit.json state and regenerating RULES_INDEX.md) and prints the report. The Quality Gate runs in structural-only mode for Branch E (no synthesis to grade — see the Quality Gate section).

---

File Formats

All file-format specifications — the File Format Template (rule files, with Scope/Load when prose guidance), the Reference File Format, the Reference Candidate Extraction strategies, and the RULES_INDEX.md Format — live in the active module's contract: {{skills_dir}}/{{self_name}}/references/module-<moduleId>.md.

Always read the contract for the resolved module and follow its formats exactly. This router intentionally does not restate them, so a new module can define its own conventions without touching this file.

Quality Gate (run before the Final Step)

A cheap, automatic self-check that runs after a branch finishes writing and before the Final Step report. It is gated by intent: run it only for ADD RULE (Branch A, including the existing-file update path), RESEARCH (Branch B), and OPTIMISE (Branch E — structural checks only, since it synthesizes no new content; see below). Skip it entirely for MIGRATE RULES (Branch C), VALIDATE INDEX (Branch D — it already exited and never reaches here), and the Step 1.5 install-from-registry short-circuit — none of those produce freshly synthesized content to grade (Migrate rephrases existing entries, Validate only reconciles the index, an install pulls a vetted upstream file). This is a checklist, not a new approval prompt: run it silently and only surface a ⚠️ line if a check fails.

Always check (every gated write — cheap structural checks):

• The rule file carries a > **Scope**: and a > **Load when**: line, both written as prose (not a bare dump of identifiers), per the module contract.
• No placeholder leak — no {Framework Name}, {Section 1}, TODO, or other template scaffolding survived into the written file.
• RULES_INDEX.md has a row for the file (new files only).

For Branch E (Optimise) these structural checks apply to the relocation: the optimised main rule still carries its > **Scope**: / > **Load when**: lines, no placeholder scaffolding leaked into the moved content or the new ## {Content} Lookup Workflow, and each created reference is wired into the main rule's > **References**: line. Reference files are not indexed, so the RULES_INDEX.md row check covers main rules only. Branch E runs none of the synthesis checks below.

Additionally — only when the branch actually synthesized content, i.e. Branch B (Research) or Branch A → A.4 with research. Do not run these on the as-is paths (A.2 cross-check-append, A.3 create-as-is, A.4 save-as-is) or Branch E (Optimise) — those write the user's exact words (or merely relocate existing text) with no synthesis, so there is no coverage or distillation to grade:

• Example coverage (#2) — the major code-facing topics the rule raises are each illustrated by an example (or explicitly waived with a stated reason), per the example inventory from research-pipeline.md B.3. Many topics raised + only one or two examples shown = fail. "Example" means whatever the active module contract defines (code snippets for code; numeric tables / worked formulas for gamedesign).
• Distilled, not copied (#8) — no long verbatim source passages; guidance is rewritten as actionable rules; any code snippets are original.
• Source Map (#6) — when the rule was built from sources, the Source Inventory was persisted into the rule file per the module contract's Source Map / provenance format.

If any check fails, fix the written file before the Final Step report (or, when a coverage gap is intentional, record the reason in the rule). Then continue to the Final Step.

Final Step: Confirm

First, run the Quality Gate above (ADD RULE / RESEARCH / OPTIMISE — structural-only for Optimise; skip for Migrate, Validate, and the install short-circuit), then produce the report below.

Reconcile .unikit.json state with disk (after Branch A / B / C / E only):

If the just-completed branch was Branch A (Add Rule), Branch B (Research), Branch C (Migrate Rules), or Branch E (Optimise) — i.e. anything that wrote a new or updated rule/reference file into .unikit/memory/<module>/ — run the CLI sync command so the generated file is registered in .unikit.json.rules.installed and RULES_INDEX.md is regenerated from authoritative state + disk:

unikit-ai rules sync

Why this matters: Branches A/B/C write files via Write/Edit directly and never touch .unikit.json. Without a sync pass, the new rule lives on disk but remains invisible to rules status, rules install <id> --force, and — most importantly — to /unikit Step 9.3, which builds its "already installed" set from .unikit.json state. Skipping sync here means the next /unikit run can try to reinstall the same rule from the registry and fail with exit 4 ("already installed"). syncRulesState Phase 1 picks up untracked .unikit/memory/**/*.md files and registers them as source: local entries; Phase 3 regenerates RULES_INDEX.md. That is exactly the reconciliation the skill needs before printing the report.

Treat a non-zero exit code as non-fatal for the skill: log the CLI output to the report and continue. The file is still on disk — the user can re-run unikit-ai rules sync manually.

Skip the sync entirely when:

• The branch was Branch D (Validate Index) — Branch D already exited with Stop. Do not proceed to any other step. and never reaches this Final Step at all.
• An install-from-registry short-circuit ran in Step 1.5 — rules install already updated state and regenerated the index, and the short-circuit terminates the skill before reaching Final Step.

Report to the user (in their configured language):

• Which module was targeted and which file was created or updated
• Summary of what was added or moved (Optimise: which sections were relocated into which reference files)
• Which branch was used (Add Rule / Research / Optimise)
• Whether Context7 enrichment was applied (Research branch only)
• Whether .unikit.json state was reconciled via rules sync (and whether the index was regenerated)
• Whether RULES_INDEX.md was updated
• If any content was redirected elsewhere (architecture, RULES.md)

Offer rules-registry promotion (after Branch A / B / C / E only):

If the just-completed branch was Branch A (Add Rule), Branch B (Research), Branch C (Migrate Rules), or Branch E (Optimise) — i.e. anything that actually wrote into .unikit/memory/<module>/ — check whether the project's registry is local:

unikit-ai rules registry show --json

Parse the JSON. If .kind == "local", ask the user whether to promote the memory changes into the registry now via AskUserQuestion:

AskUserQuestion: A local rules registry is configured at <.url from the JSON>.
Promote the memory changes into the registry now by running /unikit-rules-registry update?

Options:
1. Yes — run /unikit-rules-registry update after the Final Step report
2. No  — skip promotion, keep the changes in memory only

• Yes → finish the Final Step report first (the user must see exactly what changed in memory before promotion), then invoke unikit-rules-registry with the argument update (## Skill calls). Do not collapse the memory report into the promotion run — they are separate, ordered events.
• No → finish the Final Step report without adding any promotion hint. The user can always run /unikit-rules-registry update manually later; do not repeat the question or nag on the next invocation.

Skip the question entirely (do not ask, do not print a fallback hint) when:

• The branch was Branch D (Validate Index) — no memory writes happened.
• An install-from-registry short-circuit ran in Step 1.5 — the registry is already the source of truth for that rule.
• rules registry show --json returned a non-zero exit code or .kind is not "local".

Access Rules

Writable (this skill can create and edit, for the resolved module):

• .unikit/memory/<module>/<tier>/ — rule files in each of the module's tiers
• .unikit/memory/<module>/<tier>/references/ — supplementary reference docs, where the module contract allows them
• .unikit/memory/<module>/RULES_INDEX.md — index of all rule files in the module

Writable only during Branch C (Migrate Rules):

• .unikit/RULES.md and its topic files in .unikit/rules/ — removal of migrated entries only, plus two named edits: the <!-- @no-migrate --> tag of Keep, and the deletion of a topic file that migration emptied, together with its ## Topics row; adding new entries goes through unikit-rules skill

Read-only (this skill must NEVER modify these files):

• .unikit/config.yaml — project settings (language). Read-only for this skill.
• .unikit/system/modules.yml — module registry snapshot. Generated by unikit-ai; read-only here.
• {{skills_dir}}/{{self_name}}/references/module-<id>.md — module contracts. Read-only here.