🧠 Advanced Skills Architecture

One large skill that weaves together several phases of complex work (research → proposal → brand → operations) under a single entry point, instead of one skill per task.

It’s the leap from “a skill that does X” to “a skill that guides an end-to-end workflow”—the pattern behind systems that replace entire processes.

Single entry point · phase detection · chained or independent phases · “pushy” description to avoid under-triggering.

A chain of commands that delivers what would normally take several people: competitive research, a personalized pitch, a brand system, and an operational agent architecture.

Shows the upper limit of what a skill can deliver — and why it’s worth modeling the work as phases, not standalone tasks.

"pitch [company]" · 5 deliverables · 8 brand phases · agent team with escalation.

An embedded detector that observes what you do, identifies repeatable patterns, and offers to turn them into new skills automatically.

It's the seed of self-improvement — the same principle you see in the meta-skill (6.2) and /sessionend (6.3).

Detection triggers · “is this worth a skill?” checklist · suggestion format · create a skill when the user accepts.

Every part of the skill runs across real business fronts at the same time—it was forged through use, not on a slide.

Skills born from real problems have better descriptions, defaults, and rules than those invented "on paper."

Forged through use · explicit brand rules · cataloged high-value patterns.

A table saying "if the user says X → enter phase Y" that lets the skill route itself based on the request.

It’s what lets a huge skill have a single entry point without becoming a maze.

Request→phase map · one disambiguation question · phases that run in sequence or on their own.

Command phrases ("pitch X", "build brand book for Y", "build my agent system") that trigger the right phase and the full chain when you want everything.

You’ll leave the module knowing how to use the skill and adapt it to your brand and style.

Commands by phase · voice customization · capture your own workflow at the end.

Not a skill that performs a task, but one that governs how the others are created, gain weight, and connect — a system of skills built on top of the agent.

It's the highest-level concept in the learning path: the skill stops being content and becomes the engine for the agent's own evolution.

CLAUDE.md as the foundation · agent with name/class/level/XP · browsable skill tree.

When the agent encounters the same error twice, it creates a permanent fix; when it sees a pattern, it formalizes it as a reusable skill.

Turns failures and repetitions into capital—the agent stops rediscovering what it has already learned.

Creation loop · XP and rarity (Common→Legendary) · weights that encourage building skills.

The meta-skill analyzes its own assumptions and proposes a plan before writing a line of code—it asks what it doesn’t know instead of guessing.

It's the antidote to an agent that starts coding based on incorrect assumptions.

Gap analysis · build plan · ask before assuming.

A generated HTML page showing the agent’s growth: profile card, skill arsenal, and a graph of connections between them.

Making progress visible changes behavior—you can see what the agent knows and what’s missing.

Profile + arsenal + graph · visual levels and XP · class progression.

The adoption process: drop CLAUDE.md in the project root, open the agent, do a short onboarding, build something, and open the tree.

Shows that something this powerful can be added to a project with almost no friction.

File in the root · 30s onboarding · build → open the skill tree · optional multi-agent addon.

Where the idea can go: a skills marketplace, community profiles, agent comparisons, and merging skills that run together.

Thinking about the vision helps you design your own meta-skill with room to grow.

Marketplace · showcases/profiles · RPG comparison · skill fusion (composite skills).

Multiple agents (devs, autonomous agents, automations) keep separate context and forget everything between sessions—when one discovers something, the others never find out.

Understanding the problem is what justifies investing in a shared brain instead of another key-value store.

Isolated context · amnesia between sessions · discoveries that don't propagate.

A fact and an event are different things—each type has its own lifecycle and mutation rule: an event is immutable, a fact uses upsert, a status updates in place, and a decision preserves the why.

This taxonomy is what separates a real memory system from a pile of strings.

event (append-only) · fact (upsert by key) · status (update by subject) · decision (with reasoning).

Each memory goes through a pipeline: hash-based deduplication, supersedes chain, confidence decay, and LLM consolidation every 6 hours.

It’s what keeps the brain useful instead of turning it into a repository that only grows.

SHA-256 dedup · supersede by key/subject · 2%/day decay · consolidation that organizes things while you sleep.

No agent touches the database directly — the API validates everything, cleans credentials before storing them, and prevents one agent from deleting or changing another agent’s memory.

Shared memory between autonomous agents is only safe with a well-designed gatekeeper.

Credential scrubbing · agent isolation · timing-safe auth + rate limiting.

Every session starts with a call that returns updates from all the other agents—without you having to remember to check the logs.

It’s where coordination between agents actually happens, passively.

Briefing endpoint · excludes the agent itself · ordered by importance and recency.

A skill that, at the end of each session, gathers what happened, reflects honestly ("everything went well" is not a valid reflection), and saves a structured summary in the brain.

Closes the loop: the end of one session feeds consolidation, which informs the next briefing — each session makes all future ones smarter.

Gather → reflect → store → update · honest reflection required · a loop that compounds.