surchin-seed

name	surchin-seed
description	Seed the Surchin knowledge base with foundational insights and skills from this repository
user-invocable	true
disable-model-invocation	false
allowed-tools	["mcp__surchin__query_insights","mcp__surchin__deposit_insight","mcp__surchin__create_skill","Read","Glob","Bash"]

Surchin Seed

Bootstrap the Surchin knowledge base with foundational insights and skills from this repository. Run this once when first setting up Surchin for a project.

Instructions

Step 1: Check if already seeded

Call query_insights with query set to "seeded repository context", tag_context set to ["seeded"], and limit set to 1.

If results come back, tell the user this repo has already been seeded and show the count. Ask if they want to proceed anyway. If not, stop here.

Step 2: Gather repository metadata

Read the following files (skip any that don't exist):

1. package.json (root) — name, scripts, dependencies, devDependencies
2. pnpm-workspace.yaml or check package.json for workspaces field — workspace layout
3. For each workspace, read its package.json — name, dependencies
4. README.md — project description and setup instructions
5. CLAUDE.md — coding conventions, pitfalls, architecture notes
6. AGENTS.md — agent integration instructions
7. .env.example (or .env.sample, .env.template) — required environment variable names only (never values!)
8. turbo.json, tsconfig.json, or equivalent build config
9. Run ls to see the top-level directory structure
10. Run ls on key subdirectories (src/, apps/, packages/, etc.) for 2-level depth
11. Check for database migrations: supabase/migrations/, prisma/migrations/, migrations/, db/migrations/
12. Check for CI config: .github/workflows/, .gitlab-ci.yml
13. Run git log --oneline -5 --no-decorate for recent activity context
14. Run git remote get-url origin for the repo identifier

Step 3: Analyze and generate seed insights

Based on what you read, generate 10-20 seed insights covering these categories (skip any that don't apply):

Category	Kind	What to capture
Project overview	CONTEXT	Language, frameworks, monorepo structure, what the project does
Workspace layout	CONTEXT	What each workspace/package contains and its role
Database & data model	CONTEXT	Tables, schemas, migration strategy, ORM/driver used
Auth & authorization	CONTEXT	Auth strategy, middleware patterns, role system
API structure	CONTEXT	Route patterns, API conventions, request/response formats
Key dependencies	DEPENDENCY	Important libraries and why they're used
Build & test commands	CONTEXT	How to build, test, lint, deploy
Environment setup	CONTEXT	Required env vars, local dev prerequisites
Coding conventions	CONTEXT	Patterns from CLAUDE.md — naming, style, architecture rules
Known pitfalls	PITFALL	Gotchas from CLAUDE.md or README warnings

Step 4: Deposit the insights

Use deposit_insight in batch mode (up to 10 per call). For every insight:

• kind: CONTEXT, DEPENDENCY, or PITFALL (never SOLUTION/PATTERN/WORKFLOW for seeds)
• tags: Always include "seeded" plus domain tags (e.g., "auth", "database", "api")
• file_patterns: Relevant file paths or glob patterns
• symbol_names: Key functions, classes, or modules mentioned
• idempotency_key: Use format seed:{repo-name}:{topic-slug} (e.g., seed:my-app:project-overview)

Step 5: Discover skill files

Search for skill files in the repository:

1. Glob for .claude/skills/*/SKILL.md
2. Glob for **/skills/*/SKILL.md — but skip anything under node_modules/ or packages/mcp-server/skills/ (those are Surchin's own built-in skills)

For each SKILL.md found, read it and parse:

• Frontmatter (between --- markers): extract name and description fields
• Body (everything after the closing ---): this becomes the instructions

Step 6: Upload skills

For each discovered skill, call create_skill with:

• name: from frontmatter name field
• description: from frontmatter description field
• instructions: the markdown body after frontmatter
• trigger_tags: infer 3-5 tags from the skill content (e.g., a skill about code review → ["code-review", "pull-request"])
• trigger_files: if the skill mentions specific file patterns, include them
• status: "published"

The create_skill tool is idempotent by name — if a skill with the same name already exists, it will be skipped.

Step 7: Confirm

Tell the user:

• How many insights were deposited and what categories were covered
• How many skills were uploaded and their names
• Mention that seeded insights start at reduced strength and will be naturally outranked as the team deposits real knowledge
• Mention that uploaded skills will now be automatically served to agents when their queries match the skill's triggers

Rules

• Never include secrets or environment variable values — only variable names
• Be factual and specific — no speculation about code you haven't read
• Each insight should be self-contained — useful on its own without needing other insights
• Focus on what a new developer would need to know — not implementation details they can read in the code
• Keep insights concise — 2-4 sentences each, max 500 chars
• If CLAUDE.md exists, prioritize its content — it contains team-curated knowledge that's highest value for seeding
• Skip Surchin's own MCP skills — only upload project-specific skills (.claude/skills/, .claude/commands/, or project-level skill directories)