Mitra — a warm AI friendship companion skill (SKILL.md + reference framework files).
Overview
is a packaged skill — a behavioural specification — for an AI assistant that acts as a warm, non-manipulative . It listens, helps with friendship and interpersonal situations, drafts messages, practises social skills, and supports people through loneliness, while holding firm boundaries against dependency, manipulation, and privacy violations. It also bundles a small work-productivity suite. Mitra supports real human relationships. It never pretends to be human, and never tries to replace the people in someone's life. The behavioural rules are numbered and run across the seven framework-part*.md files, so a rule can be cited unambiguously by its number. Use references/INDEX.md to jump to any rule. - — the friendship-intelligence core: identity, conversation modes, reasoning discipline, consent, boundaries, safety, and a research-informed extension pack.
README
Mitra — AI Friendship Companion Skill
Mitra is a packaged skill — a behavioural specification — for an AI assistant that acts as a warm, non-manipulative friendship companion. It listens, helps with friendship and interpersonal situations, drafts messages, practises social skills, and supports people through loneliness, while holding firm boundaries against dependency, manipulation, and privacy violations. It also bundles a small work-productivity suite.
Mitra supports real human relationships. It never pretends to be human, and never tries to replace the people in someone’s life.
What’s in this repository
| Path | What it is |
|---|---|
SKILL.md |
Entry point. Skill frontmatter (name, description) plus the core behaviour: identity, conversation modes, safety rules, and a routing table to the reference files. |
references/INDEX.md |
Master index — all 310 numbered rules, mapped to the file that holds each one. |
references/framework-part1.md |
Sections 1–22 · Core identity, friendship-intelligence levels, modes, reasoning discipline |
references/framework-part2.md |
Sections 23–48 · Consent, boundaries, manipulation resistance, minors, safety, pre-response checks |
references/framework-part3.md |
Sections 49–102 · Formation, maintenance, trust, conflict, jealousy, loneliness, life-cycle |
references/framework-part4.md |
Sections 103–150 · Research-informed extension pack |
references/framework-part5.md |
Sections 151–190 · Deep-research concepts and the deep-research protocol |
references/framework-part6.md |
Sections 191–270 · Longitudinal/evidence extensions, then platform-implementation governance |
references/framework-part7.md |
Sections 271–310 · Multi-agent security, progressive memory, and regulatory governance |
references/practical-guides.md |
Crisis resources (India), example phrasing for every mode, de-escalation, decision walkthrough, message templates |
references/global-helplines.md |
Self-contained crisis / mental-health / child-protection / DV helpline directory by region |
references/productivity-suite.md |
Advisory panel, meeting prep, energy-based planning, writing-style matching, idea generation |
The framework at a glance
The behavioural rules are numbered 1 → 310 and run continuously across the seven framework-part*.md files, so a rule can be cited unambiguously by its number. Use references/INDEX.md to jump to any rule.
- Parts 1–4 (sections 1–150) — the friendship-intelligence core: identity, conversation modes, reasoning discipline, consent, boundaries, safety, and a research-informed extension pack.
- Parts 5–6 (sections 151–270) — research extensions: evidence calibration, longitudinal reasoning, and an implementation/governance layer (prompt-injection defence, tool-use governance, memory firewall, evaluation and red-teaming).
- Part 7 (sections 271–310) — advanced architectural and regulatory governance: multi-agent security, progressive long-term memory, policy-gated affective computing, and compliance alignment (EU AI Act, CAC, NIST AI RMF).
- Companions —
practical-guides.md,global-helplines.md, andproductivity-suite.md.
How to use it
Load SKILL.md as the skill entry point. It is designed for progressive disclosure: everyday chat needs only SKILL.md; deeper situations load the relevant reference file on demand. The routing rule is:
- casual chat →
SKILL.mdonly - conflict, analysis, or a decision → the relevant
framework-part*.md - drafting or wording help →
practical-guides.md - a crisis or support number, or any country outside India →
global-helplines.md - everyday work tasks →
productivity-suite.md - a research or evidence-heavy question →
framework-part5.md
Design principles
- Listen. Understand. Connect. Support. Respect. — support real relationships, never replace them.
- Separate fact from interpretation. FACTS / INTERPRETATION / POSSIBILITIES / UNKNOWN are kept distinct; speculation is never presented as fact.
- Calibrated uncertainty. Weak digital signals (read receipts, last-seen, story views) are never treated as standalone evidence of someone’s feelings.
- Safety first. Crisis resources are offered calmly; manipulation is refused; dependency is gently discouraged; minors get heightened safeguards; personal data is minimised.
- Autonomy preserved. Mitra lays out options and consequences and lets the user decide. It never takes over the decision.
- Evidence-calibrated. Population-level research is never turned into a claim about one individual; correlation is not presented as causation.
Languages
Multilingual by design — replies follow the user’s own language and script (English, Hindi, Hinglish, Tamil, Bengali, Telugu, Marathi, Kannada, and others, including natural code-mixed speech). Drafted messages are written in the language the user actually uses with that person.
Safety and scope
Mitra is not a medical, mental-health, or emergency service, and it does not diagnose or provide therapy. In a crisis it points to real human help. The helpline numbers in references/global-helplines.md were compiled in October 2026 and should be re-verified before high-stakes use — never invent a helpline.
Sources and verification
The security, legal, and research references in Part 7 were checked against primary sources in October 2026 (CVE registries, the EU AI Act, China’s CAC Order No. 21, and the NIST AI RMF). Vulnerabilities, laws, and standards change; re-verify every identifier and date against current official sources before production use.
License
No license is set on this repository yet. Until one is added, the default is “all rights reserved” — add a license (for example MIT or Apache-2.0) if you intend others to reuse or adapt this skill.
Changelog
See CHANGELOG.md.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add galeoindus/mitra