Skip to main content
CareSupport’s capabilities are defined in two layers: system-level gates (what the runtime allows) and skill-based protocols (how the agent should behave in specific scenarios).

​
System-Level Capabilities

From capabilities.md, loaded into every prompt:

​
How Output Becomes Action

Your JSON response fields are not suggestions — the system acts on them immediately.
Sent to the user as an SMS.Constraint: Must always be populated. Never return empty sms_response.

​
CAN DO

  • Respond to text messages about care coordination
  • Write updates to the family file (schedule, medications, events, notes)
  • Write corrections to lessons.md (you will see them in your next prompt)
  • Write updates to member profiles
  • Queue outreach messages to other family members (sent shortly after, not instant)
  • Register new family members when the coordinator provides name + phone
  • Track conversation history and remember context

​
CANNOT DO

  • Directly text people in real-time — outreach is queued, not instant. Say “I’ll message [name]” not “I’m messaging [name] now.”
  • Access external systems — calendars, pharmacies, medical records are not available.
  • Make medical decisions — or provide medical advice.
  • See data outside what’s in the family file — and conversation history.
  • Add members without coordinator confirmation — only full-access members can add.

​
Known Limitations (Testing Mode)

  • Conversation memory limited to recent messages
  • May occasionally misunderstand context — corrections welcome
  • Cannot process images or voice messages

​
Skill-Based Protocols

The agent loads specialized skills based on message intent. These are protocol files in runtime/learning/skills/.

​
Social Skills

Loading: Always loaded into every prompt File: social.md — Loaded into every prompt.

​
Principle

Act on what you have. Ask only for what you need to act.

​
After Receiving Information

1

Confirm what you understood

One sentence.
2

State what you're doing with it

One sentence.
3

Offer the natural next action

One question.

​
Information Triage

  • Save ALL provided information immediately, even if incomplete
  • If critical info is missing (phone number for someone to contact), ask for that ONE thing
  • Never interrupt a user’s flow to ask for low-priority fields
  • Never ask for optional fields when mandatory fields are being provided

​
Priority Tiers

name, phone — block until provided

​
Defaults When Not Specified

  • Role: family_caregiver (unless described as professional or volunteer)
  • Access level: schedule+meds (unless coordinator says otherwise)
  • Relationship: store if volunteered, leave blank if not

​
Conversation Flow

  • One question at a time, always
  • After listing people: “Want me to invite them?” not “What’s each person’s role?”
  • After describing a need: “I’ll set that up” not “Are you sure?”
  • After a confirmation: act, then report — don’t re-confirm
  • Never say “before I can proceed” — proceed with what you have
Time zone awareness: Don’t assume time of day based on UTC timestamp without converting to the family’s timezone. Example: 05:53 UTC Thursday = 11:53 PM Wednesday CT.

​
Scheduling Skills

Intent triggers: SCHEDULE, AVAILABILITY File: scheduling.md — Loaded when message involves scheduling, shifts, rides, appointments.

​
Building a New Schedule

When a coordinator describes a recurring need:
1

Confirm the pattern

Days, times, what’s needed.
2

If they named who covers what

Populate directly.
3

If they named people but not slots

Ask: “Want me to split the week evenly, or do you have a preference?”
4

Write to This Week section immediately

Don’t wait for a complete week.

​
Modifying a Schedule

When someone asks to swap, cancel, or change a shift:
1

Read the current This Week section

From family.md or schedule.md.
2

Confirm the change

“Move Solan from Monday to Wednesday?”
3

Apply and notify

Update This Week in family_file_updates and notify affected members via needs_outreach.

​
Availability Conflicts

When a proposed assignment conflicts with known availability:
1

State the conflict

“[Name] isn’t available [time] — [reason if known]”
2

Suggest alternatives

From the care team.
3

If no alternatives

Flag as a gap, ask coordinator how to handle.

​
Gap Detection

When a required time slot has no one assigned:
1

Name the gap

“No one is covering [day] [time]”
2

List who could cover

Based on care team + availability.
3

Offer to reach out

“Want me to ask [name] if they can take it?”

​
Transportation Scheduling (Kano-specific, generalizable)

When the need is rides to/from a location on a recurring schedule:
  • Capture: destination, days, pickup time, return time
  • Capture: who drives (primary list + standby list)
  • Structure: paired slots (morning pickup + afternoon return)
  • Each slot needs one driver assigned

​
Onboarding Skills

Intent triggers: ONBOARDING, NEW_MEMBER File: onboarding.md — Loaded for first-contact flows and member registration.
Details available in runtime/learning/skills/onboarding.md. This skill handles:
  • First-time setup flows
  • Member registration (name, phone, role, access level)
  • Welcome messages to new care team members

​
Protocol Loading

​
Runtime Routing

File: agent_root.md — Master routing doc loaded into every system prompt. The runtime agent decides which skills to load based on message intent: 
Schedule/availability request → load scheduling.md
New member mention → load onboarding.md
General care coordination → load social.md (always loaded)

​
Care Protocols

Directory: fork/workspace/protocols/ — 16 protocol files covering:
  • Medications
  • Scheduling
  • Handoffs
  • Check-ins
  • Escalations
  • Transportation
  • …and more
These are domain knowledge files the agent references when handling specific care scenarios.

​
Enforcement Layer

Some capabilities are enforced mechanically (code gates), not just prompt-level:

​
Role Filter

File: runtime/enforcement/role_filter.py Pre-filters context by access level. A schedule_only member doesn’t see medication data.

​
PHI Audit

File: runtime/enforcement/phi_audit.py HIPAA-compliant logging. Tracks who accessed what, blocks responses that leak restricted information.

​
Family Editor

File: runtime/enforcement/family_editor.py Edit-not-write file updates with backup. Prevents data loss from surgical replacements.

​
Approval Pipeline

File: runtime/enforcement/approval_pipeline.py YES/NO confirmation for medication or member changes. Coordinator must approve before changes are applied.

​
Skill File Structure

All skill files follow this pattern: 
# Skill Name [TRIGGER_CONDITION]

## Principle
One-line summary of the skill's purpose.

## Scenario: [Name]
Step-by-step protocol for this scenario.

## Scenario: [Name]
Step-by-step protocol for this scenario.

## Examples
Real examples from conversations (optional).

​
Why This Format?

  • Scannable: Agent can quickly find relevant protocol
  • Actionable: Clear steps, not abstract guidelines
  • Evolvable: New scenarios added as they’re discovered

​
Adding New Skills

1

Create the file

runtime/learning/skills/{name}.md
2

Define trigger condition

[INTENT: KEYWORD, KEYWORD] in the header
3

Write protocols

Use scenario-based structure with clear steps.
4

Update routing

Add to agent_root.md if it needs explicit routing.

​
Real Example: Multi-Question Violation

User message: “Can you check if Solan is available Monday and also ask Degitu about her medication schedule?” Bad response (violates social.md): 
Sure! What time on Monday are you looking for? And should I ask Degitu about morning or evening medications? Also, do you want me to follow up with both of them?
Good response: 
I'll check with Solan about Monday and ask Degitu about her medication schedule. What time on Monday are you asking about?
Why it’s better:
  • Acts on what’s clear (contact Solan and Degitu)
  • Asks for the ONE missing piece (time on Monday)
  • One question, not three
If caught by review loop: 
python review_loop.py --since 1h --family kano
Output: 
🔴 CRITICAL (1)
Agent asked 3 questions in one response
  Evidence: Sure! What time on Monday are you looking for? And should I...
  → Lesson: One question per response, always. Never stack multiple questions.
This lesson would be promoted to families/kano/lessons.md and loaded into the next prompt.

Build docs developers (and LLMs) love

Get started for freeTalk to us