Graphnosis Trained Skills

34 memory-trained engineering SOPs for your AI coding agent — one signed file, one double-click, loaded into a memory you own.

This repo's headline artifact is graphnosis-trained-skills.gsk — a signed, encrypted Skills Kit you open with the Graphnosis app. Inside are 34 standard operating procedures that teach an AI agent how to actually work on software: how to start a session, hunt a bug from evidence instead of guesses, ship without surprises, decide when to act alone, and recover when something goes wrong.

They're not prompt snippets or loops. They're trained skills — graphs of steps your AI walks, with goals, triggers, and failure handlers — and they're built to be retrained against your own memory so they speak your stack.

Described in the paper Borrowable Skills as Lean Un-Ganglia Subgraphs (Zenodo, 2026). Its reproducibility bundle is in route-b-artifacts/.

---

Why this is different

Most "prompt packs" are a pile of text you paste and hope. These are structured, and they plug into a memory layer:

• A skill is a graph, not a paragraph. Each one has ordered body steps connected by typed edges — seq (do in order), loop (repeat until it converges), branch (take a path on a condition), calls (invoke another skill), ctx (context that must survive) — plus 8 goal categories: Success, Out-of-scope, On-completion, Trigger, Prerequisites, On-failure, Requires, Produces. Your AI can walk the skill deterministically, capture outputs, and route to a recovery skill if a step throws.
• They compose into a practice. Skill Dispatch reads your context and routes to the right SOP; skills call sub-skills (@skill: name(args) -> $capture), loop until a smoke test passes, and branch on what they find. Autonomous Decision Authority + Self Driving Session let the agent act within boundaries it has learned you trust — and record what earns standing autonomy next time. The result is an AI that develops a practice, not just answers.
• They're memory-trained — and retrainable against yours. In Graphnosis, "training" a skill surfaces the most relevant memories from your encrypted local cortex and either appends them as grounded context (free) or rewrites the skill body with per-change attribution (Pro) — writing only to the attested, user-owned layer of your graph, never silently. These published skills were trained against an empty scope on purpose, so they ship clean and generic. Import them, then retrain against your own cortex and they start citing your architecture decisions, your conventions, your past incidents.
• Every step routes to the right model. Each step carries a @needs capability tag, so as your AI walks a skill it sends each step to the cheapest model that covers it — a local, on-device model for mechanical steps, a frontier model only where real judgment is needed.

What's inside (34 skills)

Group	Skills
Session & context	Session Start · Session End · Project Context Management · Recall Before Coding · Task Todo Management · Skill Dispatch
Build & ship	Ship Workflow · Changelog Management · Dependency Update Protocol · Generated Artifact Freshness Check · Binary Output Verification · Multi-Repo Release Coordination · Deployment Platform Ops · Feature Impact Assessment
Quality & safety	Bug Investigation · Runtime Diagnosis · Testing Cadence · Security Review Cadence · Performance Regression Check · Consistency Audit · UX Decision Gate · UX Review Checklist
Agent orchestration	Agent Delegation Decision · Subagent Management · Autonomous Decision Authority · Self Driving Session · Vibe Coding Workflow · Agentic Communication Style · MCP Tool Selection
Meta — skills about skills	Retrospective Learning · Adaptive Skill Creation · Skill Maintenance Review · Prompt Improvement · IT Architecture Review

Full per-skill detail — triggers, success criteria, outputs — is in SKILLS.md.

Import one skill, or all of them

The signed .gsk above imports the whole pack in one double-click. Each skill is also published on its own under skills/ — one signed .gsk per skill — so you can import a single SOP (say, just bug-investigation) without the rest. Same signature, same verified ✓ badge.

A taste of the kind of judgment they encode:

Runtime Diagnosis — Instrument before you theorize. Don't reason from source before you hold ground truth: pick the probe by symptom (a sampling profiler for a pegged core, devtools geometry for a stuck overlay), attribute the symptom to the right process before reading any code, design probes that split the hypothesis space, and ship both a defensive guard and the real root-cause fix — verified by re-running the same probe on the rebuilt artifact.

Ship Workflow — No surprises. Trigger words and only those count as a ship signal. Always show git status + git diff --stat and wait for an explicit go before staging. Group commits by concern. Keep security language capability-level, never vulnerability-level — the git log is public. Never push or tag without explicit confirmation.

How to load it

You need the Graphnosis desktop app (it's where the .gsk opens and where your cortex lives).

1. Download Graphnosis (macOS / Windows / Linux / Ubuntu) and unlock your cortex.
2. Double-click graphnosis-trained-skills.gsk — or in the app, open the Skills panel → Import.
3. When prompted, create/pick the "Graphnosis Trained Skills" engram as the destination.
4. You'll see a verified ✓ official badge — this pack is Ed25519-signed; if a byte were tampered with, the app would refuse it.

All 34 skills are now walkable, editable, and exportable like any skill you trained yourself. On Graphnosis Pro you can retrain any of them against your own cortex to personalize the body to your project.

Use it with Obsidian, too

Prefer to live in Obsidian? The graphnosis-obsidian-plugin connects your vault to the same cortex, so these skills and your notes share one memory.

Privacy

These skills carry zero personal data. They were trained against an empty memory scope (so nothing from a private cortex is embedded), and every body was scrubbed of project-specific paths, identities, incident details, and personal profile information before signing. What ships is generic engineering craft — verifiable by anyone, since the format is open and the pack decrypts to plain JSON.

The Graphnosis ecosystem

• graphnosis-app — the local, encrypted, deterministic memory app (desktop). Download is here.
• Graphnosis — the SDK (@nehloo/graphnosis), Apache-2.0.
• graphnosis-secure-sync — local-first encrypted storage + sync primitives.
• graphnosis-obsidian-plugin — bring your cortex into Obsidian.
• Docs: docs.graphnosis.com

License

CC BY 4.0 — use, adapt, and redistribute these skills freely, including commercially. Just attribute Graphnosis (nehloo). See CHANGELOG.md for pack versions.