reference-seek

Installation
SKILL.md

Reference Seek - Find Implementation References (v2.0.0)

Find internal patterns, quality-filtered GitHub repos with code-level insights, official docs, and external examples.

Purpose

When building a feature, find what you can reuse or learn from:

  • Internal: Existing patterns in the codebase
  • GitHub: Quality-filtered repos with actual code excerpts
  • Official Docs: context7-powered library documentation
  • External: Blog posts, tutorials, best practices

Use Cases

  • "Looking to implement OAuth login — any good references?"
  • "Find references for implementing a rate limiter"
  • "How do I implement pagination? Find me some references"
  • "Any open source projects similar to WebSocket connection management?"

Execution

Step 0: Dependency Check + Topic Classification (1 message)

Run in parallel:

0-A. gh auth check:

gh auth status
Result Variable
Authenticated GH_AVAILABLE=true
Not authenticated / gh not found GH_AVAILABLE=false

0-B. Topic extraction + classification (reasoning, same message):

Extract from user request:

  • TOPIC: Normalized English topic (e.g., "rate limiting", "OAuth authentication")
  • TOPIC_CLASS: library | pattern | concept
  • GH_QUERY: GitHub search query (e.g., rate+limiter, oauth+middleware)
  • GH_LANG_FILTER: Language filter inferred from codebase (e.g., +language:typescript)
  • CONTEXT7_LIB: Library name for context7 resolve (if library class, use directly; otherwise best guess or skip)

Examples:

  • "any rate limiter references?" → TOPIC: rate limiting, CLASS: pattern, GH_QUERY: rate+limiter+middleware, CONTEXT7_LIB: rate-limiter-flexible
  • "NextAuth implementation references" → TOPIC: NextAuth authentication, CLASS: library, GH_QUERY: nextauth+authentication, CONTEXT7_LIB: next-auth
  • "pagination references" → TOPIC: pagination implementation, CLASS: pattern, GH_QUERY: pagination+cursor, CONTEXT7_LIB: skip

Step 1: Parallel Search (1 message, up to 4 sources)

CRITICAL: Run all available tracks in one message in parallel.

Track Tool Purpose
1-A Task(Explore) Internal codebase pattern search
1-B Bash: gh api search/repositories GitHub repo search (stars-sorted, top 10)
1-C mcp__context7__resolve-library-id Library docs ID lookup
1-D WebSearch Blog/tutorial search (1 query only)

1-A. Internal Search (Explore agent):

Task(subagent_type="Explore",
     prompt="""
Find existing patterns related to [{TOPIC}] in this codebase.
Look for:
- Similar implementations or utilities
- Patterns that could be reused or extended
- Related helper functions or modules

Report as file:line format with brief description of what's reusable.
""")

1-B. GitHub Search (only if GH_AVAILABLE=true):

gh api "search/repositories?q={GH_QUERY}+{GH_LANG_FILTER}&sort=stars&per_page=10" \
  --jq '.items[] | select(.archived == false) | {
    full_name, stars: .stargazers_count, lang: .language,
    topics, license: .license.spdx_id, desc: .description,
    pushed: .pushed_at, url: .html_url, default_branch
  }'

1-C. context7 resolve (if CONTEXT7_LIB is set):

mcp__context7__resolve-library-id(libraryName="{CONTEXT7_LIB}")

1-D. WebSearch (1 query only):

WebSearch: "{TOPIC} implementation tutorial best practices 2024 2025"

When GH_AVAILABLE=false: Skip 1-B. Run WebSearch 3 times instead (v1.0 fallback):

WebSearch: "{TOPIC} implementation github"
WebSearch: "{TOPIC} open source example"
WebSearch: "{TOPIC} tutorial best practices"

Add warning at top of output: > ⚠️ GitHub API unavailable (gh not authenticated). Results from web search only.

Step 2: Quality Filter + context7 Query (sequential after Step 1)

2-A. Quality Filter (reasoning, applied to 1-B results):

Criteria Threshold
Stars >= 100 (relax to >= 50 if fewer than 2 pass)
Last push Within 24 months
License MIT, Apache-2.0, BSD-2-Clause, BSD-3-Clause, ISC, MPL-2.0
Archived Excluded (already filtered by jq)

→ Top 5 passing repos = QUALIFIED_REPOS → Top 1-2 repos = DEEP_DIVE_TARGETS

If all repos filtered out: relax stars to >= 50. If still none, take top 3 with warning badge ⚠️ Below quality threshold.

2-B. context7 docs query (only if 1-C resolve succeeded):

mcp__context7__query-docs(
  libraryId="{resolved_id}",
  query="{TOPIC} implementation patterns examples"
)

If resolve failed in 1-C, skip this step entirely.

Step 3: Repo Deep Dive (sequential after Step 2)

For each repo in DEEP_DIVE_TARGETS (1-2 repos):

3-1. File tree lookup:

gh api "repos/{owner}/{repo}/git/trees/{default_branch}?recursive=1" \
  --jq '[.tree[] | select(.type=="blob") | select(.path | test("^(src|source|lib|pkg|core)/")) | {path, size}]'

3-2. File selection (reasoning):

  • Match files by topic keywords
  • Prioritize: entry points, types/interfaces, core logic
  • Skip files > 20KB
  • Select max 3 files per repo

3-3. File content fetch:

curl -s "https://raw.githubusercontent.com/{owner}/{repo}/{default_branch}/{path}"

3-4. Code analysis (reasoning): Extract from each file:

  • Key data structures
  • Algorithm / approach
  • Configuration options
  • Notable patterns worth adopting

If file tree or fetch fails for a repo, skip its deep dive and move to the next target.

Step 4: Synthesize & Present

Combine all sources into the output format below.

Output Format

## Reference Seek: [{TOPIC}]

> Sources: Internal | GitHub ({N} repos, quality-filtered) | context7 | Web

---

### Internal References (Codebase)

#### Directly Reusable
- `{file}:{lines}` - {description}

#### Pattern Reference
- `{file}:{lines}` - {pattern description}

#### Integration Points
- `{file}:{lines}` - {connection point}

---

### GitHub References (Quality-Filtered)

1. **[{owner}/{repo}]({url})** ★{stars} | {lang} | {license} | Updated {date}
   - What: {description}
   - Topics: {topics}
   - **Key Code (`{path}`):**
     ```{lang}
     {10-30 lines of key code}
     ```
   - Takeaway: {actionable insight}

2. **[{owner}/{repo}]({url})** ★{stars} | {lang} | {license} | Updated {date}
   - What: {description}
   - Topics: {topics}
   - **Key Code (`{path}`):**
     ```{lang}
     {10-30 lines of key code}
     ```
   - Takeaway: {actionable insight}

3-5. (metadata only, no code excerpt)
3. **[{owner}/{repo}]({url})** ★{stars} | {lang} | {license} | Updated {date}
   - What: {description}

*(Filtered out: {N} repos excluded — low stars / outdated / incompatible license)*

---

### Official Docs (context7)

> Only include this section if context7 resolve succeeded in Step 1-C.

- {key concept}: {docs content}
- **Official Code Example:**
  ```{lang}
  {official example from docs}

External References (Blog / Tutorials)

  1. {title} - Key insight: {content}

Usage Suggestions

  1. Reuse: {how to leverage internal code}
  2. Reference: {most relevant repo + why}
  3. Official: {key takeaway from context7/docs}
  4. Watch out: {pitfalls or caveats}

## Error Handling

| Situation | Response |
|-----------|----------|
| gh not authenticated | Skip 1-B, use WebSearch 3x fallback, add warning at top |
| gh rate limit (403) | Fallback: `gh search repos "{GH_QUERY}" --sort stars --limit 10 --json ...` (different rate limit bucket) |
| GH results 0 | Remove language filter, retry. If still 0, show "No GitHub results found" |
| All repos filtered out | Relax stars to >= 50. If still none, show top 3 with warning badge |
| context7 resolve fails | Omit "Official Docs" section entirely |
| File tree / fetch fails | Skip that repo's deep dive, proceed to next target |
| Explore returns nothing | Show "No existing patterns found in codebase" |
| Topic too vague | Ask user for clarification before proceeding |
Weekly Installs
2
GitHub Stars
139
First Seen
Apr 8, 2026