diff --git a/.agents/skills/architecture/references/code-map.md b/.agents/skills/architecture/references/code-map.md
index 5e68dea..c2fbb48 100644
--- a/.agents/skills/architecture/references/code-map.md
+++ b/.agents/skills/architecture/references/code-map.md
@@ -13,10 +13,10 @@
 
 | Domain | Files | Top entries |
 |--------|-------|-------------|
-| src | 44 | `src/commands/doc-init.js`, `src/lib/runner.js`, `src/lib/target.js` |
+| src | 45 | `src/commands/doc-init.js`, `src/lib/runner.js`, `src/lib/target.js` |
 
 **High-churn hotspots:**
-- `src/commands/doc-init.js` - 34 changes
-- `src/commands/doc-sync.js` - 20 changes
+- `src/commands/doc-init.js` - 35 changes
+- `src/commands/doc-sync.js` - 21 changes
 - `src/lib/runner.js` - 17 changes
 
diff --git a/.agents/skills/codex-support/SKILL.md b/.agents/skills/codex-support/SKILL.md
index 0c19f40..7ba45da 100644
--- a/.agents/skills/codex-support/SKILL.md
+++ b/.agents/skills/codex-support/SKILL.md
@@ -29,6 +29,7 @@ You are working on **multi-target output support** — the system that lets aspe
 - **Target definitions:** `TARGETS.claude` (centralized) and `TARGETS.codex` (directory-scoped). Each defines paths and capability flags: `supportsHooks`, `supportsSettings`, `supportsGraph`, `supportsSkills`, `needsActivationSection`, `needsCodeMapEmbed`, `supportsMCP`. Codex also has `maxInstructionsBytes` (32 KiB) and `userSkillsDir`.
 - **Canonical generation:** Generation always produces Claude-canonical format first. Prompts always receive `CANONICAL_VARS` (hardcoded Claude paths from `doc-init.js`). Transforms run **after** generation to produce other target formats.
 - **Content transform:** `transformForTarget()` remaps paths and content. For Codex: base skill → root `AGENTS.md`, domain skills → both `.agents/skills/{domain}/SKILL.md` and source directory `AGENTS.md`. `generateCodexSkillReferences()` creates `.agents/skills/architecture/` with code-map data.
+- **Instructions file disk fallback:** `transformToDirectoryScoped` loads `instructionsFile` from disk via `repoPath` context parameter when it's not in the canonical files array (e.g., during `doc init --strategy skip-existing` or incremental `doc sync`). Uses `existsSync`/`readFileSync` from `fs`.
 - **Content sanitization:** `sanitizeCodexInstructions()` and `sanitizeCodexSkill()` strip Claude-specific references (hooks, skill-rules.json, Claude Code mentions) from Codex output.
 - **`ensureRootKeyFilesSection(content, graphSerialized)`** — Post-processes root instructions file to guarantee a `## Key Files` section with top hub files from the graph.
 - **`mergeConfiguredTargets(existing, next)`** — Merges target arrays to avoid dropping previously configured targets during narrower runs. Validates against `TARGETS` keys, deduplicates.
@@ -37,7 +38,7 @@ You are working on **multi-target output support** — the system that lets aspe
 - **Config persistence:** `.aspens.json` at repo root stores `{ targets, backend, version, saveTokens? }`. `readConfig()` returns `null` if missing **or if the config is structurally invalid**. `isValidConfig()` validates targets, backend, version, and `saveTokens` (via `isValidSaveTokensConfig()`).
 - **Feature config (`saveTokens`):** Optional object in `.aspens.json` validated by `isValidSaveTokensConfig()` — checks `enabled` (boolean), `warnAtTokens`/`compactAtTokens` (positive integers, compact > warn unless either is `MAX_SAFE_INTEGER`), `saveHandoff`/`sessionRotation` (booleans), optional `claude`/`codex` sub-objects with `enabled` and `mode`.
 - **`writeConfig` preserves feature config:** `writeConfig()` reads existing config and merges — `saveTokens` preserved unless explicitly set to `null` (intentional removal) or `undefined` (keep existing). Targets and backend also merge with existing.
-- **Multi-target publish:** `doc-sync` uses `publishFilesForTargets()` to generate output for all configured targets from a single LLM run.
+- **Multi-target publish:** `doc-sync` uses `publishFilesForTargets()` to generate output for all configured targets from a single LLM run. `repoPath` is passed through to the transform context.
 - **Codex inference tightened:** `inferConfig()` only adds `'codex'` to inferred targets when `.codex/` config dir or `.agents/skills/` dir exists.
 - **Conditional architecture ref:** Codex `buildCodexSkillRefs()` only includes the architecture skill reference when a graph was actually serialized.
 
@@ -48,9 +49,10 @@ You are working on **multi-target output support** — the system that lets aspe
 - **Codex-only restrictions:** `add agent/command/hook` and `customize agents` throw `CliError` for Codex-only repos. `add skill` works for both targets.
 - **Graph/hooks are Claude-only** — `persistGraphArtifacts()` returns data without writing files when `target.supportsGraph === false`. Hook installation skipped when `supportsHooks === false`.
 - **Config validation is defensive** — `readConfig()` treats malformed but parseable JSON (e.g., wrong types for `targets`/`backend`/`version`/`saveTokens`) as invalid and returns `null`, same as missing config.
+- **`repoPath` context is required for disk fallback** — callers of `transformForTarget` must pass `repoPath` in the context object for `instructionsFile` to load from disk when not in canonical files.
 
 ## References
 - **Patterns:** See `src/lib/target.js` for all target property definitions
 
 ---
-**Last Updated:** 2026-04-09
+**Last Updated:** 2026-04-25
diff --git a/.agents/skills/doc-sync/SKILL.md b/.agents/skills/doc-sync/SKILL.md
index 0b22073..9a4f04e 100644
--- a/.agents/skills/doc-sync/SKILL.md
+++ b/.agents/skills/doc-sync/SKILL.md
@@ -33,13 +33,13 @@ You are working on **doc-sync**, the incremental skill update command (`aspens d
 
 ## Key Concepts
 - **Monorepo-aware:** `getGitRoot(repoPath)` resolves the actual git root. `projectPrefix` (`toGitRelative`) computes the subdirectory offset. `scopeProjectFiles()` filters changed files to the project subdirectory. Diffs are fetched from `gitRoot` but file paths are project-relative.
-- **Multi-target publish:** `configuredTargets()` reads `.aspens.json` for all configured targets. `chooseSyncSourceTarget()` picks the best source (prefers Claude if both exist). LLM generates for the source target; `publishFilesForTargets()` transforms output for all other configured targets. `graphSerialized` is passed through to control conditional architecture references.
+- **Multi-target publish:** `configuredTargets()` reads `.aspens.json` for all configured targets. `chooseSyncSourceTarget()` picks the best source (prefers Claude if both exist). LLM generates for the source target; `publishFilesForTargets()` transforms output for all other configured targets. `graphSerialized` and `repoPath` are passed through to the transform context for conditional architecture references and disk-based instructions file loading.
 - **Backend routing:** `runLLM()` from `runner.js` dispatches to `runClaude()` or `runCodex()` based on `config.backend` (defaults to source target's id).
 - **Diff-based flow:** Gets `git diff HEAD~N..HEAD` from git root, scopes changed files to project prefix, then feeds diff plus existing skill contents and graph context to the selected backend.
 - **Prompt path variables:** Passes `{ skillsDir, skillFilename, instructionsFile, configDir }` from source target to `loadPrompt()` for path substitution in prompts.
 - **Refresh mode (`--refresh`):** Skips diff entirely. Reviews every skill against the current codebase. Base skill refreshed first, then domain skills in parallel batches of `PARALLEL_LIMIT` (3). Also refreshes instructions file and reports uncovered domains.
 - **Graph rebuild on every sync:** Calls `buildRepoGraph` + `persistGraphArtifacts` (with source target) to keep graph fresh. `graphSerialized` return value is captured and forwarded to `publishFilesForTargets` for conditional Codex architecture refs. Graph failure is non-fatal.
-- **Unparseable response detection:** After LLM returns, if output has content but no `<file>` tags at all, throws `CliError` instead of silently treating it as "no updates needed".
+- **Graceful response handling:** After LLM returns, if output has content but no `<file>` tags, treats it as "no updates needed" with a verbose-only warning. The prompt explicitly requests an empty response when nothing needs updating.
 - **Graph-aware skill mapping:** `mapChangesToSkills()` checks direct file matches via `fileMatchesActivation()` (from `skill-reader.js`) and also whether changed files are imported by files matching a skill's activation block.
 - **Interactive file picker:** When diff exceeds 80k chars and TTY is available, offers multiselect with skill-relevant files pre-selected.
 - **Prioritized diff:** `buildPrioritizedDiff()` gives skill-relevant files 60k char budget, everything else 20k (80k total). Cuts at `diff --git` boundaries.
@@ -53,7 +53,7 @@ You are working on **doc-sync**, the incremental skill update command (`aspens d
 ## Critical Rules
 - `runLLM` is called with `allowedTools: ['Read', 'Glob', 'Grep']` — doc-sync must never grant write tools.
 - `parseOutput` restricts paths based on `getAllowedPaths([sourceTarget])` — paths outside the allowed set are silently dropped.
-- **Unparseable output is an error** — if LLM returns text without any `<file>` tags, doc-sync throws `CliError` rather than silently proceeding with zero files.
+- **Unparseable output is a soft warning** — if LLM returns text without any `<file>` tags, doc-sync logs a verbose warning and treats it as "no updates needed" instead of throwing.
 - `getGitDiff` gracefully falls back from N commits to 1 if fewer available. `actualCommits` tracks what was used.
 - The command exits early with `CliError` if the source target's skills directory doesn't exist.
 - `checkMissingHooks()` in `bin/cli.js` only checks for Claude skills (not Codex — Codex doesn't use hooks).
@@ -64,4 +64,4 @@ You are working on **doc-sync**, the incremental skill update command (`aspens d
 - **Patterns:** `src/lib/skill-reader.js` — `GENERIC_PATH_SEGMENTS`, `fileMatchesActivation()`, `getActivationBlock()`
 
 ---
-**Last Updated:** 2026-04-08
+**Last Updated:** 2026-04-25
diff --git a/.claude/skills/codex-support/skill.md b/.claude/skills/codex-support/skill.md
index 0c19f40..7ba45da 100644
--- a/.claude/skills/codex-support/skill.md
+++ b/.claude/skills/codex-support/skill.md
@@ -29,6 +29,7 @@ You are working on **multi-target output support** — the system that lets aspe
 - **Target definitions:** `TARGETS.claude` (centralized) and `TARGETS.codex` (directory-scoped). Each defines paths and capability flags: `supportsHooks`, `supportsSettings`, `supportsGraph`, `supportsSkills`, `needsActivationSection`, `needsCodeMapEmbed`, `supportsMCP`. Codex also has `maxInstructionsBytes` (32 KiB) and `userSkillsDir`.
 - **Canonical generation:** Generation always produces Claude-canonical format first. Prompts always receive `CANONICAL_VARS` (hardcoded Claude paths from `doc-init.js`). Transforms run **after** generation to produce other target formats.
 - **Content transform:** `transformForTarget()` remaps paths and content. For Codex: base skill → root `AGENTS.md`, domain skills → both `.agents/skills/{domain}/SKILL.md` and source directory `AGENTS.md`. `generateCodexSkillReferences()` creates `.agents/skills/architecture/` with code-map data.
+- **Instructions file disk fallback:** `transformToDirectoryScoped` loads `instructionsFile` from disk via `repoPath` context parameter when it's not in the canonical files array (e.g., during `doc init --strategy skip-existing` or incremental `doc sync`). Uses `existsSync`/`readFileSync` from `fs`.
 - **Content sanitization:** `sanitizeCodexInstructions()` and `sanitizeCodexSkill()` strip Claude-specific references (hooks, skill-rules.json, Claude Code mentions) from Codex output.
 - **`ensureRootKeyFilesSection(content, graphSerialized)`** — Post-processes root instructions file to guarantee a `## Key Files` section with top hub files from the graph.
 - **`mergeConfiguredTargets(existing, next)`** — Merges target arrays to avoid dropping previously configured targets during narrower runs. Validates against `TARGETS` keys, deduplicates.
@@ -37,7 +38,7 @@ You are working on **multi-target output support** — the system that lets aspe
 - **Config persistence:** `.aspens.json` at repo root stores `{ targets, backend, version, saveTokens? }`. `readConfig()` returns `null` if missing **or if the config is structurally invalid**. `isValidConfig()` validates targets, backend, version, and `saveTokens` (via `isValidSaveTokensConfig()`).
 - **Feature config (`saveTokens`):** Optional object in `.aspens.json` validated by `isValidSaveTokensConfig()` — checks `enabled` (boolean), `warnAtTokens`/`compactAtTokens` (positive integers, compact > warn unless either is `MAX_SAFE_INTEGER`), `saveHandoff`/`sessionRotation` (booleans), optional `claude`/`codex` sub-objects with `enabled` and `mode`.
 - **`writeConfig` preserves feature config:** `writeConfig()` reads existing config and merges — `saveTokens` preserved unless explicitly set to `null` (intentional removal) or `undefined` (keep existing). Targets and backend also merge with existing.
-- **Multi-target publish:** `doc-sync` uses `publishFilesForTargets()` to generate output for all configured targets from a single LLM run.
+- **Multi-target publish:** `doc-sync` uses `publishFilesForTargets()` to generate output for all configured targets from a single LLM run. `repoPath` is passed through to the transform context.
 - **Codex inference tightened:** `inferConfig()` only adds `'codex'` to inferred targets when `.codex/` config dir or `.agents/skills/` dir exists.
 - **Conditional architecture ref:** Codex `buildCodexSkillRefs()` only includes the architecture skill reference when a graph was actually serialized.
 
@@ -48,9 +49,10 @@ You are working on **multi-target output support** — the system that lets aspe
 - **Codex-only restrictions:** `add agent/command/hook` and `customize agents` throw `CliError` for Codex-only repos. `add skill` works for both targets.
 - **Graph/hooks are Claude-only** — `persistGraphArtifacts()` returns data without writing files when `target.supportsGraph === false`. Hook installation skipped when `supportsHooks === false`.
 - **Config validation is defensive** — `readConfig()` treats malformed but parseable JSON (e.g., wrong types for `targets`/`backend`/`version`/`saveTokens`) as invalid and returns `null`, same as missing config.
+- **`repoPath` context is required for disk fallback** — callers of `transformForTarget` must pass `repoPath` in the context object for `instructionsFile` to load from disk when not in canonical files.
 
 ## References
 - **Patterns:** See `src/lib/target.js` for all target property definitions
 
 ---
-**Last Updated:** 2026-04-09
+**Last Updated:** 2026-04-25
diff --git a/.claude/skills/doc-sync/skill.md b/.claude/skills/doc-sync/skill.md
index 6ab1b4f..0242bc8 100644
--- a/.claude/skills/doc-sync/skill.md
+++ b/.claude/skills/doc-sync/skill.md
@@ -33,13 +33,13 @@ You are working on **doc-sync**, the incremental skill update command (`aspens d
 
 ## Key Concepts
 - **Monorepo-aware:** `getGitRoot(repoPath)` resolves the actual git root. `projectPrefix` (`toGitRelative`) computes the subdirectory offset. `scopeProjectFiles()` filters changed files to the project subdirectory. Diffs are fetched from `gitRoot` but file paths are project-relative.
-- **Multi-target publish:** `configuredTargets()` reads `.aspens.json` for all configured targets. `chooseSyncSourceTarget()` picks the best source (prefers Claude if both exist). LLM generates for the source target; `publishFilesForTargets()` transforms output for all other configured targets. `graphSerialized` is passed through to control conditional architecture references.
+- **Multi-target publish:** `configuredTargets()` reads `.aspens.json` for all configured targets. `chooseSyncSourceTarget()` picks the best source (prefers Claude if both exist). LLM generates for the source target; `publishFilesForTargets()` transforms output for all other configured targets. `graphSerialized` and `repoPath` are passed through to the transform context for conditional architecture references and disk-based instructions file loading.
 - **Backend routing:** `runLLM()` from `runner.js` dispatches to `runClaude()` or `runCodex()` based on `config.backend` (defaults to source target's id).
 - **Diff-based flow:** Gets `git diff HEAD~N..HEAD` from git root, scopes changed files to project prefix, then feeds diff plus existing skill contents and graph context to the selected backend.
 - **Prompt path variables:** Passes `{ skillsDir, skillFilename, instructionsFile, configDir }` from source target to `loadPrompt()` for path substitution in prompts.
 - **Refresh mode (`--refresh`):** Skips diff entirely. Reviews every skill against the current codebase. Base skill refreshed first, then domain skills in parallel batches of `PARALLEL_LIMIT` (3). Also refreshes instructions file and reports uncovered domains.
 - **Graph rebuild on every sync:** Calls `buildRepoGraph` + `persistGraphArtifacts` (with source target) to keep graph fresh. `graphSerialized` return value is captured and forwarded to `publishFilesForTargets` for conditional Codex architecture refs. Graph failure is non-fatal.
-- **Unparseable response detection:** After LLM returns, if output has content but no `<file>` tags at all, throws `CliError` instead of silently treating it as "no updates needed".
+- **Graceful response handling:** After LLM returns, if output has content but no `<file>` tags, treats it as "no updates needed" with a verbose-only warning. The prompt explicitly requests an empty response when nothing needs updating.
 - **Graph-aware skill mapping:** `mapChangesToSkills()` checks direct file matches via `fileMatchesActivation()` (from `skill-reader.js`) and also whether changed files are imported by files matching a skill's activation block.
 - **Interactive file picker:** When diff exceeds 80k chars and TTY is available, offers multiselect with skill-relevant files pre-selected.
 - **Prioritized diff:** `buildPrioritizedDiff()` gives skill-relevant files 60k char budget, everything else 20k (80k total). Cuts at `diff --git` boundaries.
@@ -53,7 +53,7 @@ You are working on **doc-sync**, the incremental skill update command (`aspens d
 ## Critical Rules
 - `runLLM` is called with `allowedTools: ['Read', 'Glob', 'Grep']` — doc-sync must never grant write tools.
 - `parseOutput` restricts paths based on `getAllowedPaths([sourceTarget])` — paths outside the allowed set are silently dropped.
-- **Unparseable output is an error** — if LLM returns text without any `<file>` tags, doc-sync throws `CliError` rather than silently proceeding with zero files.
+- **Unparseable output is a soft warning** — if LLM returns text without any `<file>` tags, doc-sync logs a verbose warning and treats it as "no updates needed" instead of throwing.
 - `getGitDiff` gracefully falls back from N commits to 1 if fewer available. `actualCommits` tracks what was used.
 - The command exits early with `CliError` if the source target's skills directory doesn't exist.
 - `checkMissingHooks()` in `bin/cli.js` only checks for Claude skills (not Codex — Codex doesn't use hooks).
@@ -64,4 +64,4 @@ You are working on **doc-sync**, the incremental skill update command (`aspens d
 - **Patterns:** `src/lib/skill-reader.js` — `GENERIC_PATH_SEGMENTS`, `fileMatchesActivation()`, `getActivationBlock()`
 
 ---
-**Last Updated:** 2026-04-08
+**Last Updated:** 2026-04-25
diff --git a/AGENTS.md b/AGENTS.md
index 145a909..fe6fde2 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -1,3 +1,41 @@
+# aspens
+
+## Skills
+
+- `.agents/skills/doc-sync/SKILL.md` — Incremental skill updater that maps git diffs to affected skills and optionally auto-syncs via a post-commit hook
+- `.agents/skills/codex-support/SKILL.md` — Multi-target output system — target abstraction, backend routing, content transforms for Codex CLI and future targets
+- `.agents/skills/architecture/SKILL.md` — Import graph and code-map reference for structural changes.
+
+## Commands
+
+- `npm test` — run Vitest (`vitest run`)
+- `npm start` — run the CLI (`node bin/cli.js`)
+- `npm run lint` — no-op check (`echo 'No linter configured yet' && exit 0`)
+- `aspens scan [path]` — deterministic repo scan
+- `aspens doc init [path]` — generate skills, hooks, and instructions file (`--target claude|codex|all`, `--recommended` for full recommended setup including save-tokens, agents, and doc-sync hook)
+- `aspens doc impact [path]` — show freshness, coverage, drift, and LLM interpretation of generated context (interactive apply for repairs)
+- `aspens doc sync [path]` — update docs from recent diffs
+- `aspens doc graph [path]` — rebuild `.agents/skills/architecture/references/code-map.md`
+- `aspens add <type> [name]` — install bundled templates
+- `aspens save-tokens [path]` — install token-saving session settings (`--recommended`, `--remove`)
+
+## Release
+
+- Release workflow: `/Users/MV/aspenkit/dev/release.md`
+
+## Conventions
+
+- ESM only: use `import`/`export`; never `require()`.
+- Prefer `CliError` from command handlers; top-level handling lives in `bin/cli.js`.
+- `es-module-lexer` must be initialized before `parse()`.
+- Keep target/backend semantics straight: target is output format/location; backend is the generating CLI. Persist config in `.aspens.json`.
+- Do not duplicate base-skill guidance here; consult `.agents/skills/base/SKILL.md` for deeper repo context.
+
+## Behavior
+
+- **Verify before claiming** — Never state that something is configured, running, scheduled, or complete without confirming it first. If you haven't verified it in this session, say so rather than assuming.
+- **Make sure code is running** — If you suggest code changes, ensure the code is running and tested before claiming the task is done.
+
 ## Key Files
 
 **Hub files (most depended-on):**
@@ -11,10 +49,10 @@
 
 | Domain | Files | Top entries |
 |--------|-------|-------------|
-| src | 44 | `src/commands/doc-init.js`, `src/lib/runner.js`, `src/lib/target.js` |
+| src | 45 | `src/commands/doc-init.js`, `src/lib/runner.js`, `src/lib/target.js` |
 
 **High-churn hotspots:**
-- `src/commands/doc-init.js` - 34 changes
-- `src/commands/doc-sync.js` - 20 changes
+- `src/commands/doc-init.js` - 35 changes
+- `src/commands/doc-sync.js` - 21 changes
 - `src/lib/runner.js` - 17 changes
 
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1fb02f3..bd7aa90 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,6 +1,11 @@
 # Changelog
 
-## [Unreleased]
+## [0.7.3] - 2026-04-25
+
+### Fixed
+- **`doc sync` crash on unstructured LLM response** — when the LLM responds with explanatory text instead of `<file>` tags, `doc sync` now treats it as "no updates needed" instead of throwing. Also tightened the prompt to request an empty response when nothing needs updating.
+- **Codex `AGENTS.md` missing most content** — the codex transform produced only hub files and domain clusters (~17 lines) instead of the full instructions derived from `CLAUDE.md`. The transform now loads `CLAUDE.md` from disk when it's not in the canonical files (e.g., during `doc init --strategy skip-existing` or incremental `doc sync`).
+
 
 ## [0.7.2] - 2026-04-16
 
diff --git a/README.md b/README.md
index 58fa0ac..dc60549 100644
--- a/README.md
+++ b/README.md
@@ -7,7 +7,7 @@
 **Your CLAUDE.md stopped working. Here's why.**
 
 [![npm version](https://img.shields.io/npm/v/aspens.svg)](https://www.npmjs.com/package/aspens)
-[![npm downloads](https://img.shields.io/npm/dm/aspens.svg)](https://www.npmjs.com/package/aspens)
+[![npm downloads](https://img.shields.io/npm/dy/aspens.svg)](https://www.npmjs.com/package/aspens)
 [![GitHub stars](https://img.shields.io/github/stars/aspenkit/aspens)](https://github.com/aspenkit/aspens)
 [![MIT License](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
 
@@ -19,14 +19,28 @@ aspens replaces the monolith with scoped skill files (~35 lines each) generated
 
 Works with Claude Code, Codex, or both.
 
+## Install
+
 ```bash
-npx aspens doc init --recommended
+npm install -g aspens
+```
+
+Then in your project:
+
+```bash
+aspens doc init --recommended
 ```
 
-Then verify what it generated:
+Verify what it generated:
 
 ```bash
-npx aspens doc impact
+aspens doc impact
+```
+
+Or run without installing:
+
+```bash
+npx aspens doc init --recommended
 ```
 
 ![aspens demo](demo/demo-full.gif)
diff --git a/package-lock.json b/package-lock.json
index 3156e16..0096ad1 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -1,12 +1,12 @@
 {
   "name": "aspens",
-  "version": "0.7.2",
+  "version": "0.7.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "aspens",
-      "version": "0.7.2",
+      "version": "0.7.3",
       "hasInstallScript": true,
       "license": "MIT",
       "dependencies": {
diff --git a/package.json b/package.json
index 6782f18..e20a11a 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "aspens",
-  "version": "0.7.2",
+  "version": "0.7.3",
   "description": "Keep coding-agent context accurate as your codebase changes",
   "type": "module",
   "bin": {
diff --git a/src/commands/doc-init.js b/src/commands/doc-init.js
index c796e72..1e3a618 100644
--- a/src/commands/doc-init.js
+++ b/src/commands/doc-init.js
@@ -137,7 +137,7 @@ function validateGeneratedChunk(files, repoPath) {
   return files;
 }
 
-function buildOutputFilesForTargets(canonicalFiles, targets, scan, graphSerialized) {
+function buildOutputFilesForTargets(canonicalFiles, targets, scan, graphSerialized, repoPath) {
   let outputFiles = [...canonicalFiles];
   const nonClaudeTargets = targets.filter(target => target.id !== 'claude');
 
@@ -146,6 +146,7 @@ function buildOutputFilesForTargets(canonicalFiles, targets, scan, graphSerializ
       const transformed = transformForTarget(canonicalFiles, TARGETS.claude, target, {
         scanResult: scan,
         graphSerialized,
+        repoPath,
       });
 
       const transformValidation = validateTransformedFiles(transformed);
@@ -716,7 +717,7 @@ export async function docInitCommand(path, options) {
   // For --target all: keep canonical + add transformed for each non-Claude target.
   const canonicalFiles = [...allFiles]; // preserve originals
   if (!shouldWriteIncrementally) {
-    allFiles = buildOutputFilesForTargets(canonicalFiles, targets, scan, graphSerialized);
+    allFiles = buildOutputFilesForTargets(canonicalFiles, targets, scan, graphSerialized, repoPath);
   }
 
   let results = [];
@@ -1546,7 +1547,7 @@ async function generateChunked(repoPath, scan, repoGraph, domains, baseOnly, tim
       if (writeIncrementally && files.length > 0) {
         writeIncrementalOutputs(
           repoPath,
-          buildOutputFilesForTargets(allFiles, targets, scan, graphSerialized),
+          buildOutputFilesForTargets(allFiles, targets, scan, graphSerialized, repoPath),
           shouldForce,
           writeState
         );
@@ -1628,7 +1629,7 @@ async function generateChunked(repoPath, scan, repoGraph, domains, baseOnly, tim
             if (writeIncrementally) {
               writeIncrementalOutputs(
                 repoPath,
-                buildOutputFilesForTargets(allFiles, targets, scan, graphSerialized),
+                buildOutputFilesForTargets(allFiles, targets, scan, graphSerialized, repoPath),
                 shouldForce,
                 writeState
               );
@@ -1711,7 +1712,7 @@ async function generateChunked(repoPath, scan, repoGraph, domains, baseOnly, tim
         if (writeIncrementally && files.length > 0) {
           writeIncrementalOutputs(
             repoPath,
-            buildOutputFilesForTargets(allFiles, targets, scan, graphSerialized),
+            buildOutputFilesForTargets(allFiles, targets, scan, graphSerialized, repoPath),
             shouldForce,
             writeState
           );
diff --git a/src/commands/doc-sync.js b/src/commands/doc-sync.js
index f02a4c7..011b644 100644
--- a/src/commands/doc-sync.js
+++ b/src/commands/doc-sync.js
@@ -62,7 +62,7 @@ function chooseSyncSourceTarget(repoPath, targets) {
   return targets[0] || TARGETS.claude;
 }
 
-function publishFilesForTargets(baseFiles, sourceTarget, publishTargets, scan, graphSerialized = null) {
+function publishFilesForTargets(baseFiles, sourceTarget, publishTargets, scan, graphSerialized = null, repoPath = null) {
   const published = [];
 
   for (const target of publishTargets) {
@@ -74,6 +74,7 @@ function publishFilesForTargets(baseFiles, sourceTarget, publishTargets, scan, g
     const transformed = transformForTarget(baseFiles, sourceTarget, target, {
       scanResult: scan,
       graphSerialized,
+      repoPath,
     });
     published.push(...transformed);
   }
@@ -308,12 +309,12 @@ ${truncate(instructionsContent, 5000)}
 
   // Step 6: Parse output
   const baseFiles = parseOutput(result.text, allowedPaths);
-  const hasFileTags = /<file\s+path=/i.test(result.text);
-  if (baseFiles.length === 0 && result.text.trim().length > 0 && !hasFileTags) {
-    syncSpinner.stop(pc.red('Unparseable response'));
-    throw new CliError('LLM returned content without <file path="..."> tags. Aborting instead of treating it as "no updates needed".');
+  if (baseFiles.length === 0 && result.text.trim().length > 0 && !/<file\s+path=/i.test(result.text)) {
+    if (verbose) {
+      p.log.warn('LLM responded without <file> tags — treating as no updates needed.');
+    }
   }
-  const files = publishFilesForTargets(baseFiles, sourceTarget, publishTargets, scan, graphSerialized);
+  const files = publishFilesForTargets(baseFiles, sourceTarget, publishTargets, scan, graphSerialized, repoPath);
 
   if (files.length === 0) {
     syncSpinner.stop('No updates needed');
@@ -646,7 +647,7 @@ async function refreshAllSkills(repoPath, options, sourceTarget, publishTargets
     return;
   }
 
-  const filesToWrite = publishFilesForTargets(allUpdatedFiles, sourceTarget, publishTargets, scan, graphSerialized);
+  const filesToWrite = publishFilesForTargets(allUpdatedFiles, sourceTarget, publishTargets, scan, graphSerialized, repoPath);
   const directWriteFiles = filesToWrite.filter(f => !(f.path.endsWith('/AGENTS.md') && f.path !== 'AGENTS.md'));
   const dirScopedFiles = filesToWrite.filter(f => f.path.endsWith('/AGENTS.md') && f.path !== 'AGENTS.md');
   const results = [
diff --git a/src/lib/target-transform.js b/src/lib/target-transform.js
index e95045a..194ecef 100644
--- a/src/lib/target-transform.js
+++ b/src/lib/target-transform.js
@@ -6,6 +6,7 @@
  */
 
 import { join } from 'path';
+import { readFileSync } from 'fs';
 
 export function transformForTarget(files, sourceTarget, destTarget, context) {
   if (sourceTarget.id === destTarget.id) return files;
@@ -44,11 +45,19 @@ function remapCentralizedPath(filePath, sourceTarget, destTarget) {
 function transformToDirectoryScoped(files, sourceTarget, destTarget, context) {
   const scanResult = context?.scanResult;
   const graphSerialized = context?.graphSerialized;
+  const repoPath = context?.repoPath;
   const result = [];
 
   const baseSkillPrefix = sourceTarget.skillsDir + '/base/';
   const baseSkill = files.find(file => file.path.startsWith(baseSkillPrefix));
-  const instructionsFile = files.find(file => file.path === sourceTarget.instructionsFile);
+  let instructionsFile = files.find(file => file.path === sourceTarget.instructionsFile);
+
+  if (!instructionsFile && repoPath && sourceTarget.instructionsFile) {
+    try {
+      const content = readFileSync(join(repoPath, sourceTarget.instructionsFile), 'utf8');
+      instructionsFile = { path: sourceTarget.instructionsFile, content };
+    } catch {}
+  }
   const domainSkills = files.filter(file =>
     file !== baseSkill &&
     file !== instructionsFile &&
diff --git a/src/prompts/doc-sync.md b/src/prompts/doc-sync.md
index d1fc4f6..4235d38 100644
--- a/src/prompts/doc-sync.md
+++ b/src/prompts/doc-sync.md
@@ -22,7 +22,7 @@ Return ONLY the files that need updating, wrapped in XML tags:
 - Only output files that actually changed. If a skill doesn't need updates, don't include it.
 - Output the COMPLETE file content, not a diff or patch. The file will be written as-is.
 - Use `<file path="...">` and `</file>` tags exactly as shown.
-- If nothing needs updating (cosmetic changes, test-only changes, docs-only changes), output nothing.
+- If nothing needs updating (cosmetic changes, test-only changes, docs-only changes), output an empty response — no text at all, not even an explanation.
 
 ## Rules