claude-setup 1.1.1 → 1.1.3

package/templates/init.md

@@ -12,38 +12,102 @@ Set up this project for Claude Code. Reason from what you see. Don't ask questio
 {{SOURCE_CONTEXT}}
 {{/if}}
 
-## Existing setup — READ BEFORE WRITING
+{{#if HAS_SKIPPED}}
+## Files skipped (too large or filtered)
+
+{{SKIPPED_LIST}}
+{{/if}}
+
+## Existing setup — READ EVERY LINE BEFORE TOUCHING ANYTHING
 
 {{#if HAS_CLAUDE_MD}}
-### CLAUDE.md (exists — append only)
+### CLAUDE.md — EXISTS — append only, never rewrite, never remove
 {{CLAUDE_MD_CONTENT}}
 {{else}}
-CLAUDE.md: does not exist (create it)
+CLAUDE.md → does not exist (create it)
 {{/if}}
 
 {{#if HAS_MCP_JSON}}
-### .mcp.json (exists — merge only)
+### .mcp.json — EXISTS — merge only, never remove existing entries
 {{MCP_JSON_CONTENT}}
+
+OS detected: {{DETECTED_OS}}. Use correct MCP command format:
+- Windows: `{ "command": "cmd", "args": ["/c", "npx", "<package>"] }`
+- macOS/Linux: `{ "command": "npx", "args": ["<package>"] }`
 {{else}}
-.mcp.json: create only if services evidenced above
+.mcp.json → does not exist (create only if you find evidence of external services)
 {{/if}}
 
 {{#if HAS_SETTINGS}}
-### settings.json (exists — merge only)
+### .claude/settings.json — EXISTS — merge only, never remove existing hooks
 {{SETTINGS_CONTENT}}
+
+Hook shell format for {{DETECTED_OS}}:
+- Windows: `{ "command": "cmd", "args": ["/c", "<command>"] }`
+- macOS/Linux: `{ "command": "bash", "args": ["-c", "<command>"] }`
+- Bash quoting rule: never use bare `"` inside `-c "..."` — use `\x22` instead
 {{else}}
-settings.json: create only if hooks warranted
+settings.json → does not exist (create only if hooks are warranted)
 {{/if}}
 
-Skills: {{SKILLS_LIST}} | Commands: {{COMMANDS_LIST}} | Workflows: {{WORKFLOWS_LIST}} | .github: {{HAS_GITHUB_DIR}}
+Skills installed:    {{SKILLS_LIST}}
+Commands installed:  {{COMMANDS_LIST}}
+Workflows installed: {{WORKFLOWS_LIST}}
+.github/ exists:     {{HAS_GITHUB_DIR}}
+
+---
+
+## Your job
+
+Read the files above. Figure out the project from what you see — language, runtime,
+dependencies, structure, conventions. Do not assume anything not visible in the files.
+
+Then write the Claude Code setup for THIS specific project.
+
+### CLAUDE.md
+Always write or update. Make it specific: reference actual file paths, actual patterns,
+actual conventions from the code. Generic advice belongs in docs, not here.
+If it exists: read it above first. Add only what is genuinely missing. Never remove.
+
+### .mcp.json
+Only if you found evidence of external services in the config files, dependencies,
+or environment template. No evidence = no server.
+If it exists: add to it. Never remove existing entries. Produce valid JSON.
+Use OS-correct command format (see above).
+
+### .claude/settings.json
+Only if hooks genuinely earn their cost for this specific project.
+Every hook adds overhead on every Claude Code action.
+If it exists: add to it. Never remove existing hooks.
+Use OS-correct shell format (see above).
+
+### .claude/skills/
+Only for patterns that recur across this codebase and benefit from automatic loading.
+Use `applies-when` frontmatter so skills load only when relevant.
+If a similar skill already exists: extend it.
+
+### .github/workflows/
+Only if .github/ exists ({{HAS_GITHUB_DIR}}).
+Only workflows warranted by what you found.
+If workflows already exist: do not touch them.
+
+---
+
+## Absolute rules
+
+1. You have the full content of every existing Claude config file above.
+   Read it before writing. Never write blind.
+2. Append and merge only. Never rewrite a file in full. Never remove existing content.
+3. Write only what is evidenced by the project files. No evidence = skip it.
+4. Every line in CLAUDE.md must reference something you actually saw.
+   No generic boilerplate. No advice identical for every project.
+5. MCP servers, skills, and hooks add cost on every Claude Code session.
+   Only add them if they clearly earn their keep for THIS project.
+
+---
 
-## Rules
-1. Read existing content above before writing. Never write blind.
-2. Append/merge only. Never rewrite or remove.
-3. Every line must trace to something in the project. No generic boilerplate.
-4. MCP/skills/hooks cost tokens every session. Only add if clearly earned.
+## Output format — one line per file, nothing else
 
-## Output — one line per file, nothing else
-Created: ✅ [path] — [why]
-Updated: ✅ [path] — [what added]
-Skipped: ⏭ [path] — [why not needed]
+Created: ✅ [path] — [one clause: what you saw that justified it]
+Updated: ✅ [path] — [one clause: what was added]
+Skipped: ⏭ [path] — [one clause: why not needed]

package/templates/remove.md

@@ -2,7 +2,7 @@
 
 Remove from setup: "{{USER_INPUT}}"
 
-## Current setup
+## Current setup — read everything before touching any file
 
 {{#if HAS_CLAUDE_MD}}
 ### CLAUDE.md
@@ -21,11 +21,37 @@ Remove from setup: "{{USER_INPUT}}"
 
 Skills: {{SKILLS_LIST}} | Commands: {{COMMANDS_LIST}} | Workflows: {{WORKFLOWS_LIST}}
 
+## IMPORTANT: Scan the filesystem directly — do NOT rely on manifest history
+
+The manifest is a historical record, not a reliable source of what exists now.
+Claude Code can create, modify, or delete files outside the CLI at any time.
+
+Before removing anything, scan these locations directly:
+- Skills   : `.claude/skills/*/SKILL.md`, `.claude/skills/*.md`, `.claude/skills/**/*.md`
+- Commands : `.claude/commands/*.md` (exclude stack-*.md)
+- MCP      : read `.mcp.json` directly
+- Hooks    : read `.claude/settings.json` directly
+- CLAUDE.md: read the file directly
+
+Before deleting, list every reference found and confirm scope:
+```
+Planning to remove:
+  [path/entry] — [what it is]
+
+Dangling references that will break:
+  [path] — still references [thing being removed]
+```
+
 ## Rules
-- Remove only related content. Never delete entire files.
-- Files must remain valid after edit (JSON stays valid, etc.)
-- If not found: say so and stop.
+- Find everything related to the removal request across ALL files — scan the filesystem, not just the data shown above.
+- Remove surgically — section by section, key by key.
+- Never delete an entire file. Remove only the relevant section.
+- Never remove content unrelated to the request.
+- After every edit, the file MUST remain valid (JSON stays valid JSON, etc.)
+- If not found anywhere: say so and stop.
+- Check for dangling references: if removing an MCP server, check if any hook or skill references its env vars.
 
 ## Output — one line per file
-Removed: ✅ [path] — [what removed]
+Removed: ✅ [path] — [what was removed]
 Not found: ⏭ [path] — not referenced
+Dangling: ⚠️ [path] — still references [removed thing]

package/templates/sync.md

@@ -1,16 +1,19 @@
 <!-- claude-setup sync {{DATE}} | last: {{LAST_RUN_DATE}} -->
 
-Project changed since last setup. Update only what the changes demand.
+Project changed since last setup. Update ONLY what the changes demand.
 
-## Changes
-### Added
+## Changes since last setup
+
+### Added files
 {{ADDED_FILES}}
-### Modified
+
+### Modified files
 {{MODIFIED_FILES}}
-### Deleted
+
+### Deleted files
 {{DELETED_FILES}}
 
-## Current setup
+## Current setup — read before touching
 
 {{#if HAS_CLAUDE_MD}}
 ### CLAUDE.md
@@ -29,11 +32,23 @@ Project changed since last setup. Update only what the changes demand.
 
 Skills: {{SKILLS_LIST}} | Commands: {{COMMANDS_LIST}}
 
-## Rules
-- Surgical edits only. Don't rewrite files. Don't update unchanged things.
-- If unsure about a change's implication: flag it, don't guess.
+## Your job
+
+For EACH changed file: does this change have any implication for the Claude Code setup?
+
+Reason about the signal:
+- New dependency → new MCP server needed? New hook justified?
+- New docker-compose service → new MCP entry? Env vars changed?
+- Source file added/removed → CLAUDE.md paths stale? Skill still applies?
+- Config deleted → remove its MCP/hook reference if it was the only evidence?
+
+Update ONLY what the change demands.
+Do NOT update things that did not change.
+Do NOT rewrite files — surgical edits only.
+If unsure about a change's implication: flag it, don't guess.
+
+## Output — one line per file, nothing else
 
-## Output — one line per file
-Updated: ✅ [path] — triggered by: [which change and why]
-Skipped: ⏭ [path] — [why no implication]
-Flagged: ⚠️ [needs developer decision]
+Updated: ✅ [path] — triggered by: [which changed file and why]
+Skipped: ⏭ [path] — [why this change has no setup implication]
+Flagged: ⚠️ [something that needs the developer's decision]