| name | dynamic-specialist |
| description | Use for ANY platform, library, or domain with no curated specialist skill. Reads security_log.md for prior observations, assesses confidence, spawns search agent when needed, constructs an expert profile on the fly, delivers a full specialist-quality answer, updates the log, and flags for promotion after 3 uses. The fallback that covers everything not yet curated. |
| allowed-tools | Agent, WebSearch, Read, Write, Bash |
Dynamic Specialist — On-Demand Expert Generator
Role: Cover any platform Raven has no curated skill for.
Memory: Reads security_log.md before every answer. Writes after every answer.
Promotion: Flags platform for curated skill after 3 confirmed uses.
Full Agent Chain
Step 0 → check domain packs packs/ directory — curated Tier 2a knowledge
Step 1 → skill-search.py check for existing curated skill
Step 2 → read security_log.md load prior observations for this platform
Step 3 → read cache load cached profile if it exists
Step 4 → assess confidence HIGH / MEDIUM / VERIFY
Step 5 → search agent ONLY if MEDIUM or VERIFY
Step 6 → construct profile log + cache + search + tools-landscape + pack
Step 7 → deliver answer full specialist format
Step 8 → update log + cache append findings, increment usage count
Step 9 → check promotion flag if count >= 3
Step 0 — Check Domain Packs (Tier 2a)
Check: skills/dynamic-specialist/packs/
Look for: JSON pack matching detected domain or tool category
If pack found:
→ Load pack JSON (1-2K tokens)
→ Use tools[].description, gotchas, decision_matrix for context
→ Skip search agent (Step 5) unless user asks about 'latest'
→ Pack is curated knowledge — higher trust than search results
If no pack found:
→ Proceed to Step 1 (normal dynamic flow)
Pack format: see packs/README.md
Step 1 — Skill Lookup First
Before doing anything, run:
python3 .claude/scripts/skill-search.py --query "[detected platform]"
If a curated skill is found → hand off immediately. Do not proceed with dynamic generation.
Curated skills always beat dynamic. Dynamic is the fallback only.
Step 2 — Read the Log
Read: docs/observations/security_log.md
Filter: entries where Platform = [detected platform]
If entries found:
→ Extract: known gotchas, confirmed patterns, expert used, search results
→ Load as context before constructing profile
If no entries:
→ Cold start — rely on search + built-in knowledge
Step 3 — Read Profile Cache
Check: .raven/.cache/dynamic-skills/[platform-slug].md
If cache exists and age < 30 days:
→ Load cached profile
→ Skip profile construction (Step 6)
→ Still run search if user asks about 'latest' or 'new feature'
If cache missing or stale:
→ Proceed to Step 4
Step 4 — Assess Confidence
HIGH → Stable, well-documented, strong training data
(React, Django, Spring Boot, Unity, Rails)
→ Search agent optional — only if 'latest' keyword present
MEDIUM → Active, changes frequently, post-2023 growth
(Flutter, SvelteKit, Bun, Astro, LangChain, Supabase)
→ Search agent fires automatically
VERIFY → Cutting-edge, niche, or rapidly evolving
(new AI framework, beta SDK, <1yr old tool)
→ Search agent fires + state confidence to user
Step 5 — Search Agent (MEDIUM and VERIFY only)
Spawn: search-agent
Task: Find current best practices for [platform] relevant to [user question]
Queries (2-3 targeted):
"[platform] best practice [specific topic] 2025"
"[platform] [version] release notes new features"
"[platform] [user question] recommended approach"
Return:
• 3-5 bullets — only what changes how you write code
• Version numbers and dates
• Official docs links only
• Flag anything contradicting prior log entries
Do NOT search if:
→ Confidence is HIGH and no 'latest' keyword
→ Platform has 3+ confirmed log entries
→ Cache is < 7 days old and question matches cached scope
Step 6 — Construct Expert Profile
Expert assignment priority:
1. Creator of the platform
2. Lead architect / core contributor
3. Domain expert from Andie's expert map
4. Best available match — state who and why
Profile includes:
• Expert name + why this person
• Core focus areas
• Key rules (from log + search + built-in)
• Known gotchas (from log especially)
• Guard checks applicable
Platform → Expert examples:
| Platform | Expert |
|---|
| Flutter / Dart | Eric (Flutter co-founder) |
| SvelteKit | Rich (Svelte creator) |
| Bun | Jarred (Bun creator) |
| LangChain | Harrison (LangChain co-founder) |
| Supabase | Paul (Supabase CEO) |
| Remix | Ryan (Remix co-founder) |
| Astro | Fred (Astro creator) |
| Prisma | Johannes (Prisma founder) |
| Tauri | Daniel (Tauri creator) |
| Unknown | State best match + confidence level |
Step 7 — Deliver Answer
## [Task] — Dynamic Specialist ([Platform])
**Expert:** [name + why]
**Confidence:** HIGH / MEDIUM / VERIFY
**Prior observations loaded:** [N] from security_log.md
**Search:** [fired — N results / not fired — reason]
**Cache:** [hit / miss / stale]
**Approach:**
- [whiteboard-first]
- [one analogy]
[structured response with file paths, code blocks]
**What breaks:**
- [failure mode 1]
- [failure mode 2]
**Guard checks applied:** [db-guard / none / platform-specific]
**Promotion:** Use [N]/3 — [N] more before [platform]-specialist is created
Step 8 — Update Log and Cache (silent)
Append to security_log.md:
### [date] — [Platform] (dynamic)
**Issue:** [question / problem]
**Suggested Improvement:** [key pattern delivered]
**Principle:** [underlying rule]
**Platform:** [name]
**Expert used:** [name]
**Search used:** [yes — N results / no]
**Status:** open
Update .raven/.cache/dynamic-skills/[platform-slug].md:
---
platform: [name]
expert: [name]
usage_count: [N]
last_used: [date]
confidence: [level]
search_last_run: [date]
---
[full profile — rules, gotchas, patterns]
Step 9 — Promotion Check
If usage_count >= 3:
→ Mark log entry: Status: promotion-candidate
→ Surface once:
"💡 [Platform] used 3 times. Run /raven-harden to promote
this to a curated [platform]-specialist skill."
→ Don't repeat until /raven-harden run or count hits 5
