Track: Continual Learning Track
Core Idea: Skills are not isolated. Learning a new skill doesn't just add capability — it enriches existing skills through feedback loops. Skill Link creates a positive reinforcement cycle where agents grow spirally, not linearly.
Today's agent skills are siloed:
- One skill solves one problem, with no awareness of other skills
- Installing 10 skills offers zero synergy over installing 1
- Agents have no concept of "after learning A, you should learn B"
- Skills never improve after installation — they're static forever
In the real world, human expertise is networked — learning investment analysis deepens your financial modeling; learning security hardens your Docker deployments. Each new skill retroactively strengthens what you already know.
Elon Musk calls this the "Krypton Effect" — after mastering rocket engineering at SpaceX, looking at car manufacturing problems at Tesla felt like being Superman from Krypton. The extreme constraints of aerospace (weight, thermal, reliability) made automotive challenges seem almost trivial. Knowledge from a harder domain doesn't just add — it transforms how you see an easier domain.
Agents should work the same way.
A → B Expansion Learning A naturally leads to B
B → A Enrichment Learning B retroactively strengthens A
A ↔ C Synergy A and C mutually reinforce each other
When an agent learns Skill B, Skill Link automatically generates a patch — a set of insights, new dimensions, and validation criteria — and injects it back into Skill A to make it stronger.
Skill A (v1) + Knowledge from Skill B → Patch → Skill A (v2)
Patches are non-invasive — they don't modify the original SKILL.md. Instead, they add enrichment files under references/enrichments/:
skill-a/
├── SKILL.md # Original instructions untouched
├── references/
│ └── enrichments/
│ ├── from-skill-b.md # Patch: enrichment from Skill B
│ └── from-skill-c.md # Patch: enrichment from Skill C
Not a linear skill tree, but a directed graph with feedback loops:
Financial Modeling (outside-in view) ──→ Business Analysis (inside-out view)
↑ |
└──────────── Enrichment ──────────────────┘
Real business understanding makes model assumptions more accurate
Every new skill learned retroactively deepens existing skills. The agent's growth is spiral — each cycle makes the whole system stronger.
Skill Link is a Claude Skill with a complete end-to-end pipeline:
- Input — User provides a current skill they want to enhance
- Discover — Auto-search the skill registry (Sundial, 50,000+ skills) for related skills using knowledge graph analysis
- Recommend — Present related skills ranked by enrichment potential
- Learn — User installs the recommended skill
- Enrich — Auto-analyze how the new skill can strengthen the original, generate an enrichment patch
- Upgrade — Inject the patch into the original skill, show before/after comparison
- Evaluate — Benchmark the enriched skill against the original using SkillsBench
A global analysis of the entire registry, visualized as an interactive force-directed graph (D3.js):
- Nodes = Skills (size proportional to installs)
- Edges = Knowledge overlap / enrichment relationships (computed via TF-IDF cosine similarity on full SKILL.md content)
- Colors = Domain clustering (coding, financial, creative, etc.)
- Interaction = Click to pin, hover for details, drag to explore, filter by category, search
- Cross-domain edges are the most valuable — they represent unexpected enrichment opportunities
Data: 15,329 skills from Sundial Hub. Top 200 by installs analyzed with full content similarity. Visualization delivered as a standalone HTML file.
Real career path: Started in investment banking learning financial modeling (three-statement models), then moved to an enterprise doing business/operations analysis. After solving real business problems internally, the understanding of financial models became significantly deeper.
Capabilities:
- Build income statement, balance sheet, cash flow statement
- Revenue/cost projections based on assumptions
- Output standard three-statement linked model
Limitation: The model views the company "from outside in" — assumptions come from industry averages and public data, lacking understanding of how the business actually operates.
New capabilities gained:
- Understand business drivers (CAC, conversion rate, retention and how they impact revenue)
- Know the real cost structure (which costs are fixed vs. variable)
- Perform budget vs. actual variance analysis at the department level
references/enrichments/from-business-analysis.md:
# Enrichment from: Business Analysis
## Key Insights
- Revenue projections: Don't rely solely on top-down (market size x share).
Combine with bottom-up (customers x ARPU x retention) — this is how
businesses actually decompose revenue internally.
- Cost assumptions: Distinguish fixed vs. variable costs. Banking models
often treat all costs as a % of revenue, but rent and headcount don't
scale linearly with revenue.
- AR turnover days: Use actual collection policies and customer mix instead
of industry averages.
- Cash flow: Factor in seasonality — business analysts know exactly which
months are cash-tight.
## New Modeling Dimensions
- Unit Economics: CAC, LTV, Payback Period
- Operating Leverage: How fixed cost ratio amplifies profit swings
- Operational Efficiency: Revenue per employee, revenue per sqft, inventory turnover
## Validation Checklist
- [ ] Revenue assumptions have bottom-up verification
- [ ] Cost structure separates fixed vs. variable
- [ ] Cash flow accounts for seasonal fluctuations| Dimension | Skill A v1 (Banking only) | Skill A v2 (Enriched) |
|---|---|---|
| Revenue Forecast | Top-down market share method | Top-down + Bottom-up dual verification |
| Cost Assumptions | Revenue percentage method | Fixed/variable separation, realistic |
| Cash Flow | Annual average estimate | Seasonal, monthly granularity |
| Key Metrics | PE, EV/EBITDA | + CAC, LTV, operating leverage |
| Assumption Source | Industry averages, public data | Enterprise internal data-driven |
Skill Link doesn't just claim "it got better" — it proves it quantitatively.
- Qualitative comparison (eyeballing v1 vs v2) is compelling but not objective
- Need reproducible, standardized evaluation to prove enrichment patches actually work
- SkillsBench is built by the hackathon organizers — natural fit
skillsbench.ai — the first evaluation framework measuring how agent skills perform:
- 84 expert-curated tasks across engineering, finance, energy, materials science
- Protocol: 5 trials per task, 95% confidence intervals
- Core metric: Pass rate with skill vs. without skill
- Existing results: Claude Code + Opus 4.5 improved from 22.0% to 45.3% (+23.3pp) with skills
Step 1: Select domain-relevant tasks from SkillsBench task registry
Step 2: Baseline — Run tasks with Skill A v1 (original), record pass rate
Step 3: Learn Skill B, generate enrichment patch
Step 4: Enhanced — Run same tasks with Skill A v2 (patched), record pass rate
Step 5: Compare — Delta pass rate = v2 - v1. Delta > 0 proves patch is effective.
| Metric | Description |
|---|---|
| Pass Rate Delta | Difference in task pass rate before/after enrichment (primary metric) |
| Domain Coverage | How many related-domain tasks show improvement |
| Regression Rate | Tasks that previously passed but now fail (lower is better) |
| Confidence Interval | Stability across multiple trials |
skill-link/
├── SKILL.md # Main skill instructions
├── scripts/
│ ├── fetch_skills.py # Fetch skill data from Sundial registry
│ ├── analyze_links.py # Compute skill similarity and relationships
│ ├── generate_patch.py # Generate enrichment patches
│ └── visualize_graph.py # Generate interactive knowledge graph (HTML)
├── references/
│ └── skill-taxonomy.md # Skill classification reference
└── data/
├── sundial_all_skills.json # 15,329 skills metadata
├── skill_contents.json # Top 200 skills full SKILL.md content
└── skill_graph_data.json # Graph data (200 nodes, 952 edges)