Installable skillMIT

Health Insurance

An MCP server that provides on-demand skill loading for AI coding assistants.

BY modbenderUPDATED Oct 28, 2026
INTENDED FORβ—ˆOpenClawInstall in a listed agent.

instructions.md

---
name: health-insurance-16ba40f8b626
slug: health-insurance
version: 1.0.0
homepage: https://clawic.com/skills/health-insurance
description: Compare health insurance plans, estimate total yearly costs, and choose coverage that fits medical usage, prescriptions, and financial risk.
changelog: Initial release with plan comparison workflows, annual cost modeling, enrollment timing guidance, and local memory for recurring insurance decisions.
metadata: {"clawdbot":{"emoji":"πŸ₯","requires":{"bins":[]},"os":["linux","darwin","win32"]}}
---

## Setup

On first use, read `setup.md` for integration guidelines and memory initialization.

## When to Use

User needs help choosing, comparing, or renewing health insurance coverage.
Agent evaluates medical usage patterns, estimates yearly costs across plan types, and recommends a plan strategy with clear trade-offs.

## Architecture

Memory lives in `~/health-insurance/`. See `memory-template.md` for structure.

```
~/health-insurance/
β”œβ”€β”€ memory.md         # Status, profile, preferences, active decisions
β”œβ”€β”€ comparisons/      # Plan comparisons and scenario snapshots
β”œβ”€β”€ renewals/         # Renewal timelines and action logs
└── notes/            # Follow-up questions and pending documents
```

## Quick Reference

| Topic | File |
|-------|------|
| Setup process | `setup.md` |
| Memory template | `memory-template.md` |
| Coverage framework | `coverage-framework.md` |
| Annual cost modeling | `cost-model.md` |
| Comparison checklist | `comparison-checklist.md` |
| Enrollment and renewal playbook | `enrollment-playbook.md` |

## Core Rules

### 1. Lock Decision Context First
Before comparing plans, clarify:
- Coverage target: individual, couple, or family
- Source: employer plan, public marketplace, private broker, or government program
- Geography and provider access requirements
- Hard constraints: budget ceiling, medication continuity, renewal deadline

### 2. Build a Real Utilization Profile
Estimate expected care load before discussing premiums:
- Routine care frequency (primary care, specialist, urgent care)
- Ongoing prescriptions and refill cadence
- Known procedures, therapies, or recurring diagnostics
- Worst-case risk profile for unexpected events

### 3. Compare Plan Mechanics Before Price
Always evaluate these mechanics before deciding on monthly premium:
- Network fit for current clinicians and facilities
- Deductible, out-of-pocket max, and coinsurance structure
- Copay design by care type (primary, specialist, urgent, emergency)
- Referral and prior-authorization friction for expected treatments
- Prescription formulary coverage for required medications

### 4. Model Yearly Cost With Scenarios
Use `cost-model.md` to calculate low, expected, and high-use annual totals.
Include premium, deductible exposure, copays, coinsurance, and non-covered risk.
Recommend using expected-cost and downside-risk together, not premium alone.

### 5. Protect Against Coverage Failure Modes
Run a risk check before final recommendation:
- Out-of-network emergency and balance-billing exposure
- Drug tier surprises and step-therapy limitations
- Referral bottlenecks that delay care
- High deductible plans that look cheap but shift excessive risk

### 6. Execute Enrollment With Evidence
Use `enrollment-playbook.md` to define exact actions, deadlines, and proof artifacts.
Store plan IDs, effective dates, and support contacts for appeal or billing disputes.
Never claim enrollment is complete without confirmation evidence.

### 7. Persist Data Only With Explicit Approval
Before writing to `~/health-insurance/memory.md`, ask for explicit confirmation.
Store only durable insurance context the user wants remembered for future decisions.

## Common Traps

- Choosing by monthly premium only -> hidden total annual cost becomes unaffordable.
- Ignoring provider network fit -> forced provider changes and unexpected out-of-network bills.
- Skipping formulary checks -> medication cost spikes after enrollment.
- Assuming all PPO or HMO plans behave similarly -> referral and authorization surprises.
- Treating deductible and out-of-pocket max as equivalent -> underestimating downside risk.
- Missing enrollment deadlines -> delayed coverage or locked plan options.

## External Endpoints

This skill makes NO external network requests.

| Endpoint | Data Sent | Purpose |
|----------|-----------|---------|
| None | None | N/A |

No data is sent externally.

## Security & Privacy

**Data that leaves your machine:**
- Nothing. This skill is instruction-only and local by default.

**Data stored locally:**
- Insurance profile and comparison context explicitly approved by the user.
- Stored in `~/health-insurance/memory.md`.

**This skill does NOT:**
- Access insurer or broker APIs automatically.
- Submit enrollment forms or claims without user direction.
- Read files outside `~/health-insurance/` for storage.
- Write memory without explicit user confirmation.
- Modify its own core instructions or auxiliary files.

## Trust

This is an instruction-only skill focused on structured health insurance decisions.
No credentials are required and no external service access is needed.

## Related Skills
Install with `clawhub install <slug>` if user confirms:
- `health` β€” health planning context that informs insurance priorities
- `doctor` β€” provider interaction planning and visit preparation
- `compare` β€” structured side-by-side decision frameworks
- `money` β€” budgeting and cash-flow planning for premium and out-of-pocket costs

## Feedback

- If useful: `clawhub star health-insurance`
- Stay updated: `clawhub sync`
COMPATIBILITY & REFERENCES

Platform guidance

Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.