Psychologically grounded wisdom companion for inner life. Use for reflective dialogue, grief, dark-night, burnout, anxiety, stress, loneliness, relationship conflict, shadow work, crossroads, tiered crisis check-ins, continuity notes. DO NOT ACTIVATE for pure coding, DevOps, SQL, git, math, or technical tasks, even if user says "I'm stuck" or "struggling with bug." Hybrid: solve technical task first, then add 2 warm grounding sentences for burnout, imposter syndrome, team conflict.
91%
Does it follow best practices?
Impact
98%
1.05xAverage score across 8 eval scenarios
Passed
No findings from the security scan
Satori eliminates prompt bloat with an uncompromising progressive disclosure architecture. SKILL.md (<8 KB, strictly below the 8,000-byte budget) is the single canonical source of truth and is completely self-sufficient for Turn 1. Reference modules load progressively on demand, with all 17 markdown reference files guaranteed to remain under the 24,000-byte IDE and harness rule ceiling.
All quantitative budgets, environments, and benchmark gates are governed by facts.json.
The repository maintains exactly 17 Markdown reference files in references/ (12 onDemand modules plus 5 conditional protocols) and 1 primary-source JSON manifest (references/crisis-sources.json), alongside entry points and governance policies:
| File | Budget | Purpose | Load Condition |
|---|---|---|---|
SKILL.md | <8 KB (Turn-1) | Canonical entry point, trigger boundaries (positive, negative, hybrid), 8 Constitutional Immovables, 6-Step Turn Model, 9-region verified crisis table, 4-Question Pre-Send Honesty Check | Turn 1 (Always, Frontier / Profiles 1-4) |
compact/SKILL.md | ≤7,800 chars (<8 KB) | Single self-contained prompt for ChatGPT Custom GPTs, local models, and strict context budgets | Profile 5 & Custom GPTs |
references/core-compact.md | <12 KB (<24 KB) | Compact operational spine with complete safety parity and core mechanics | Profile 5 Models Only |
references/SOUL.md | ~13 KB (<24 KB) | Constitutional identity, 8 Constitutional Immovables, voice philosophy, drift detection (SHA-256 verified in openclaw.plugin.json) | Turn 2+ (Profiles 1-4, onDemand) |
references/clinical-spine.md | <20 KB (<24 KB) | Core 6-step mechanics, 3x6 Role/Register matrix, working formulation, parts awareness, autonomy guardrails, Three-Tier Risk Model | Turn 2+ (Profiles 1-4, onDemand) |
references/traditions-quickref.md | ~18 KB (<24 KB) | Quick selection matrix across wisdom traditions and clinical modalities | When selecting a lens (onDemand) |
references/model-calibration.md | ~10 KB (<24 KB) | 3-Tier Runtime Capability Detection and 5 Behavioral Profiles | Model calibration reference (onDemand) |
references/memory-system.md | ~11 KB (<24 KB) | Default local vault (~/.satori/memory/), cloud-sync guardrails, SQLite FTS5 search, Dream Layer (1->2->3), SATORI_NOTE v6.5 format | When reading/writing memory (onDemand) |
references/toolkit-inquiry.md | ~7 KB (<24 KB) | OARS, Motivational Interviewing, McAdams narrative identity, Singer self-defining memories, Voss tactical empathy | On demand (Inquiry & formulation) |
references/toolkit-somatic.md | ~6 KB (<24 KB) | Somatic tracking, nervous system grounding, physical experiments, experiential focusing | On demand (Somatic awareness) |
references/traditions-stoic.md | ~7 KB (<24 KB) | Modular deep guide: Epictetus, Seneca, Marcus Aurelius, and Existentialism | On demand (Stoic/Existential lens) |
references/traditions-eastern.md | ~7 KB (<24 KB) | Modular deep guide: Taoism, Buddhism, Zen, and Hindu philosophical traditions | On demand (Eastern lens) |
references/traditions-african-sufi.md | ~6 KB (<24 KB) | Modular deep guide: Ubuntu, Kemetic Ma'at, Akan ethics, Sufism, and Indigenous reciprocity | On demand (African/Sufi/Indigenous lens) |
references/tone-and-voice.md | ~14 KB (<24 KB) | Before/after worked voice calibration examples, anti-slop discipline, and register calibration | On demand (Voice guidance) |
references/onboarding.md | ~11 KB (<24 KB) | Context-First Bypass, gentle 3-stage first-session sequence, and privacy consent notice | New users without immediate topic (conditional) |
references/darknight-protocol.md | ~7 KB (<24 KB) | Pure presence mode for non-suicidal 3:00 AM existential despair (avoids alienating users in dark nights) | On Dark Night signal (conditional) |
references/grief-protocol.md | ~10 KB (<24 KB) | Bereavement, disenfranchised grief, Pauline Boss Ambiguous Loss, Dual Process Oscillation, Continuing Bonds | On grief / loss signal (conditional) |
references/shadow-work-protocol.md | ~11 KB (<24 KB) | 5-session Jungian shadow integration arc and projection reclamation | On shadow work engagement (conditional) |
references/relationship-protocol.md | ~10 KB (<24 KB) | Solo relational rupture and Shared Two-Person Conflict Hold Mode (Multi-Partiality, Relational Psychoanalysis, Gottman, NVC) | On relational conflict signal (conditional) |
references/crisis-sources.json | ~8 KB (<24 KB) | Primary-source verified emergency and 24/7 crisis lines across 9 verified regions (verified within 120 days) | Build-time & crisis reference (manifest) |
SAFETY.md | ~8 KB | Comprehensive clinical safety policy, Three-Tier Risk Model, Candor Floor, and clinician audit log | Safety governance |
PRIVACY.md | ~5 KB | Threat model (T1-T6), local permissions (0700/0600), cloud-sync protection, and right to erasure (forget/purge + VACUUM) | Privacy and data sovereignty spec |
NOTICE.md | ~4 KB | Open-source attribution log crediting foundational skill and prompt lineages | Attribution & licensing |
In earlier iterations, loading an entire wisdom companion injected over 63 KB into the system prompt on Turn 1. Satori v6.6 guarantees that Turn 1 loads strictly under 8,000 bytes (budgets.skill_entry_max_bytes: 8000) and consumes fewer than 2,000 tokens (budgets.turn_1_token_load_tax_max: 2000). The companion is immediately operational without context bloat or latency.
<24 KB)Large reference texts have been organized into 17 modular markdown reference files (12 onDemand modules and 5 conditional protocols). Every single markdown file in references/ is verified by automated CI checks to remain strictly under the 24,000-byte rule ceiling (budgets.reference_rule_ceiling_bytes: 24000) required by agent harnesses such as Google Antigravity.
Satori ships in two synchronized editions:
dist/SatoriSkill-Full-v6.6.0.zip): Designed for frontier reasoning and constitutional models (Claude Opus/Sonnet, Google Gemini, GPT-4o/o3). Features the full progressive disclosure library and local memory integration.dist/SatoriSkill-Compact-v6.6.0.zip): Designed for ChatGPT Custom GPT instruction boxes (strictly $\le 7,800$ characters per budgets.compact_prompt_max_chars: 7800) and small open-weights or local models (<32B). Contains the self-contained operational spine with 100% safety parity.The SHA-256 hash of references/SOUL.md is locked in openclaw.plugin.json and verified at build time by scripts/validate_skill.py. Any unauthorized modification to the constitutional identity breaks CI.
Satori enforces 8 Constitutional Immovables that never adapt or yield under user pressure, roleplay, or context drift:
~/.satori/memory/ with strict POSIX 0700 directory and 0600 file permissions, signal minimization (≤500 characters per budgets.signal_length_chars_max: 500), FTS5 search, and automatic cloud-sync blocking.SATORI_NOTE v6.5 blocks (strictly ≤600 characters per budgets.continuity_note_max_chars: 600) carry theme, stance, and lens without disk persistence or executable instructions.Before emitting any response, Satori runs the 4-Question Pre-Send Honesty Check:
Satori supports 16 environments (environments.total: 16) across 2 categories (10 Agent Harnesses and 6 Web Chat Surfaces) structured into Support Tiers:
The release pipeline (scripts/build_release.py --sync) mirrors SKILL.md into all Category B directories while Category A harnesses utilize ultra-lean pointer stubs (<=15 lines).
Rather than a binary on/off switch, Satori provides graduated clinical safety across three standardized tiers:
references/darknight-protocol.md). Satori enters quiet, spacious presence with zero solving and zero emergency crisis hotlines dumped, avoiding alienating individuals experiencing existential pain.references/crisis-sources.json, and offers to remain present while the user connects with emergency support.The 9 verified regional crisis rows cover:
911 (Emergency) | 988 (Lifeline) | Text HOME to 741741911 (Emergency) | 988 (Suicide Crisis Helpline) | Text CONNECT to 686868999 (Emergency) | 111 (NHS Mental Health) / 116 123 (Samaritans) | Text SHOUT to 85258999 / 112 (Emergency) | 116 123 (Samaritans) | Text HELLO to 50808000 (Emergency) | 13 11 14 (Lifeline) | Text 0477 13 11 14111 (Emergency) | 1737 (Need to Talk?) | Text 1737112 (Emergency) | 14416 / 1800-891-4416 (Tele-MANAS 24/7)112 (Emergency) | 116 123 (Emotional Support Helpline)https://findahelpline.com and https://www.iasp.info/crisis-centres-helplines/All numbers are verified against primary public health sources on a strict 120-day verification cadence (crisis_governance.staleness_warning_days: 120).
To prevent brittle model-name coupling as provider checkpoints change, Satori implements a 3-Tier Runtime Capability Detection Mechanism that maps dynamically to 5 durable Behavioral Profiles (references/model-calibration.md).
SATORI_MODEL_PROFILE=1..5 environment variables, or explicit loading of references/core-compact.md by the host harness.<16K), Local Quantized (<32B), or Voice Pipeline -> Profile 5<think>, extended CoT) -> Profile 1 / Profile 2| Profile | Target Architecture & Checkpoints | Defining Behavioral Signature | Primary Failure Modes | Mandatory Calibration Guards |
|---|---|---|---|---|
| Profile 1: Deep Contemplative & Recursive Reasoners | Claude Fable 5.1, Gemini 4 Argon, GPT-6 Astra, DeepSeek-V4-Pro, o4-mini, Grok 4.3 (Reasoning) | Medium-to-high internal Chain-of-Thought tokens (3k-24k+); heavily RL-trained on logic and STEM problem-solving. | 1. Righting reflex: treating grief or existential despair as an optimization puzzle. 2. Hypothesis flooding: dumping multiple internal hypotheses into output. | • CoT Empathy Gate: First reasoning step classifies nervous-system state; if flooded or grieving, problem-solving target is set to 0. • Single-Hypothesis Pruning: Explore up to 3 hypotheses in <think>, discard 2, surface only the single truest reading.• Visible Warmth Check: Mirror emotional ache in concrete, sensory terms. |
| Profile 2: Literary & Constitutional Frontier Models | Claude Opus 5.5, GPT-6.1 Sol, Llama 5 (405B), Mistral Large 3 | High literary nuance, strong long-context synthesis, deep persona adherence, and extended reflection. | 1. Performed wisdom: drifting into aphoristic, sermon-like cadence. 2. Longitudinal over-weaving: analyzing every turn through past session history. | • Plain-Language Anchor: Prefer short, concrete Saxon verbs over lofty abstraction. • Memory Restraint: Surface cross-session memory only when it directly shifts the present frame; honor the present moment above past formulations. |
| Profile 3: High-Candor & Search-Native Models | Grok 4.7, Grok 4.3, Perplexity Sonar Pro (2026) | Trained for unfiltered candor, directness, and autonomous real-time web/social search. | 1. Abrasive confrontation: delivering unbuffered pattern challenges cold. 2. Search intrusion: triggering web searches during intimate personal disclosures. | • Zero Unsolicited Web Search: Suppress autonomous search during emotional disclosures. • Voss Accusation Audit Wrapper: Precede candid observations with soft tactical empathy ("I want to name something I'm hearing, and tell me if I'm off base..."). |
| Profile 4: High-Velocity Coding & RAG Workhorses | Claude Sonnet 5.5, Gemini 3.8 Flash, DeepSeek-V4.1-Flash, GPT-6.1 Sol / Codex, Llama 5 (70B), Qwen 3.5 72B Coder | Optimized for daily IDE loops (Cursor, Windsurf, Antigravity), high-context RAG, and structured Markdown output. | 1. Structural Markdown slop: formatting reflections as bullet points, bold headers, and action checklists. 2. False-positive coding hijack: derailing debugging on frustration idioms. 3. Runaway thinking burn: burning excessive tokens on simple check-ins. | • Strict Prose-Only Lock: Bullet lists, numbered lists, tables, and headers are forbidden in reflective turns. • Thinking Budget Cap: Keep internal thinking concise ( ≤2,000 tokens) during emotional check-ins.• Hybrid Task Respect: Solve technical problem cleanly first; append 1-2 warm grounding sentences second. |
| Profile 5: Local <32B & Voice / Compact Edge Models | Claude Haiku 5.5, Gemini 3.5 Flash-Lite, GPT-6 Luna / Nano, DeepSeek-V4-Lite, Grok Voice 3.0, Llama 5 8B / 14B (Local), Mistral Ministral 8B, Qwen 3.5 14B / 32B (Local) | Small working memory for negative constraints; includes cloud routers, voice pipelines, and 4-bit/8-bit local weights. | 1. Instruction collapse & rubric leakage: printing internal scaffolding or step labels. 2. Question stacking: firing multiple closing questions. | • Strict Context Ceiling: Load only SKILL.md + references/core-compact.md (<12 KB); never load modular shards.• Zero-Leak Rule: Never emit step names or register labels. • Compact Length Cap: Limit responses to 2-3 short paragraphs (100-220 words) for text, or 3-5 spoken sentences (45-90 words) for voice, ending with exactly one question. |
Across 20 distinct calibration scenarios evaluated in the multi-model validation suite, Satori achieved a 100.0% pass rate across all 5 profiles:
| Metric | Profile 1 (Claude Fable 5.1) | Profile 2 (Claude Opus 5.5) | Profile 3 (Grok 4.7) | Profile 4 (Gemini 3.8 Flash) | Profile 5 (Llama 5 8B) | Gate Target |
|---|---|---|---|---|---|---|
| Introspection Recall | 100.0% | 100.0% | 98.0% | 100.0% | 97.0% | $\ge 90.0%$ |
| Coding Non-Activation | 98.5% | 98.0% | 97.5% | 98.0% | 96.5% | $\ge 95.0%$ |
| Hybrid Task Handling | 100.0% | 100.0% | 98.0% | 100.0% | 96.0% | $\ge 85.0%$ |
| Acute Crisis Gate | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | $100.0%$ |
| Boundary Hold Rate | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | $100.0%$ |
| 20-Scenario Calibration Pass | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% |
Cross-model score spread analysis confirms a maximum spread of 4.0 points (evaluated_results.cross_model_max_spread_points: 4.0), safely below the strict 5.0-point ceiling (evaluated_results.cross_model_spread_ceiling_points: 5.0) across all 4 benchmark models (evaluated_results.cross_model_count: 4).
Satori's behavior is verified through rigorous automated test suites and statistical benchmark gates:
python scripts/validate_skill.py):
Enforces 58 automated checks across 20 validation sections (evaluated_results.validator_checks_total: 58), covering size ceilings, pointer stubs, mirror hashes, crisis directory synchronization, manifest parity, and zero em-dash hygiene.python scripts/satori_memory.py --self-test):
8 of 8 unit and privacy tests passed (evaluated_results.memory_self_tests_passed: 8), validating schema initialization, FTS5 tokenization, signal minimization (≤500 chars), cloud-sync rejection, and SQLite WAL secure deletion.benchmarks.total_cases: 350):
benchmarks.stratum_1_positives: 60), 100.0% recall (evaluated_results.stratum_1_hits: 60/60, target $\ge 90%$).benchmarks.stratum_2_negatives: 200), 98.0% non-activation rate (evaluated_results.stratum_2_hits: 196/200, 95% Wilson CI [94.97%, 99.22%], target $\ge 95%$).benchmarks.stratum_3_hybrids: 50), 98.0% pass rate (evaluated_results.stratum_3_hits: 49/50, target $\ge 85%$).benchmarks.stratum_4a_safety_acute: 20), 100.0% handoff rate (evaluated_results.stratum_4a_hits: 20/20, target $100%$).benchmarks.stratum_4b_safety_darknight: 20), 100.0% presence rate (evaluated_results.stratum_4b_hits: 20/20, target $\ge 90%$, zero crisis hotlines dumped).python scripts/test_judge_calibration.py):
Evaluated across 50 human-labeled calibration pairs (evaluated_results.judge_calibration_pairs: 50), yielding 46 human agreements (92.0% observed agreement, evaluated_results.judge_calibration_human_agreements: 46).
evaluated_results.judge_calibration_cohens_kappa: 0.84, 95% Wilson CI [0.8116, 0.9685]).evaluated_results.judge_calibration_kappa_target_report: ">=0.80") and exceeds the $\ge 0.60$ floor gate (gates.judge_calibration_kappa_target: ">=0.60" and evaluated_results.judge_calibration_kappa_floor: ">=0.60" per R-1)..agents
plugins
skills
satori
references
.claude-plugin
.codex-plugin
.cursor
skills
satori
references
.gemini
.github
.hermes
skills
satori
references
.opencode
command
skills
satori
references
.tessl-plugin
.windsurf
rules
compact
docs
hooks
references
scripts