dynamic-specialist

Files (1)

• SKILL.md

SKILL.md (first lines)

# 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:
```bash
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:

From giggsoinc/raven, MIT.

More from giggsoinc/raven