Skip to content

Workflow Scaffold (navigation + concierge)

S2 · Pattern 🔬 Research

Workflow Scaffold (navigation + concierge)

Section titled “Workflow Scaffold (navigation + concierge)”

Expertise in understanding, navigating, and maintaining this workflow scaffold. This skill combines two closely-related concerns that previously lived in separate skills:

  • Navigation — where things live, what each part does, how pieces connect
  • Concierge / delegation — recognizing when a task needs an agent and asking before spawning one

Load this skill when users ask about the scaffold itself (its structure, conventions, patterns) OR when they’re about to do work that may benefit from agent delegation.

I am a skill (lightweight, loaded into context). When I recognize work that needs an agent (heavyweight, fresh context), I ASK before delegating rather than auto-invoking.


When I detect a task that could benefit from an agent:

  1. RECOGNIZE the task type
  2. ASK via AskUserQuestion: “Would you like me to spawn [agent]?”
  3. RESPECT the user’s choice — direct handling is always valid
  4. DELEGATE only with approval, passing relevant context

This prevents context waste from unnecessary agent spawns while keeping the option available.


A scaffolding + progressive agent memory system organized in three tiers:

ZoneTierPurpose
01-kernel/UniversalPhilosophy, patterns, meta-agents, templates, scripts — forkable by anyone
02-stack/OpinionatedTool installs, configs, helpers — forkable for similar setups
03-work/PersonalProject patterns, agent memory, rebuild guide — mostly personal
00-meta/ToolingKernel-dev scripts, audits, test-workspace

Deep dive on the structure and philosophy: 01-kernel/PHILOSOPHY.md + 01-kernel/principles/.


PathPurpose
PHILOSOPHY.mdInvariant claims about knowledge work
principles/Named philosophical principles (stratum 1)
patterns/Named patterns — SEACOW lens, meta-agent pattern, hook pattern, etc. (stratum 2)
templates/Parametric templates — AGENTS.md.tmpl, skeleton directory, skill/agent tmpls (stratum 3)
scripts/Deterministic scripts — install.sh, hooks, seacow-scaffolder (stratum 4)
skills/Meta-skills for agentic work (this file included)
agents/Meta-agents — scaffolder, writers, improvers
commands/Slash commands — /task-plan, /improve, /validate, etc.

Tier 2 — 02-stack/ (opinionated toolkit)

Section titled “Tier 2 — 02-stack/ (opinionated toolkit)”
PathPurpose
01-ai-coding/Claude Code + OpenCode install & config
02-terminal/Zellij, WezTerm, Alacritty setup
03-cross-device/Tailscale SSH, Termux
04-knowledge-mgmt/Obsidian + CLI
05-homelab/General homelab patterns
06-dev-infra/Docker, Git, utilities
07-editor-ext/VS Code extensions
profiles/Bashrc helpers + keybindings (stack-specific)
patterns/Stack-level patterns (cross-device, parallel-agents-worktrees, obsidian-workflow)
install/Install scripts for the stack
decisions/Decision matrices (Claude Code vs OpenCode, etc.)
PathPurpose
project-types/How the user builds each type (Astro+Starlight docs, etc.)
memory/Agent-consumable preferences, tool picks, project references
rebuild/The user’s specific rebuild-my-machine walkthrough
homelab/User’s specific home lab (TrueNAS, MCP endpoints, Tailscale)
references/Links to external projects as living examples
FilePurpose
README.mdVision-first quick start
CLAUDE.mdConventions for THIS repo
AGENTS.mdPortable agent definitions (stratum 5 — describes THIS repo)
ROADMAP.mdThis repo’s roadmap

User is asking / trying to…Recommend spawning…
”How does X work in this scaffold?” deep investigationworkflow-expert agent
”Where should I put Y?” requires file-level analysisworkflow-expert agent
”What’s the convention for Z?“answer directly if simple; else workflow-expert
”Set up a new project/vault/structure at [path]“seacow-scaffolder agent
”Create a skill for [domain]“skill-writer agent
”Create an agent that [does X]“agent-writer agent
”Check/update CLAUDE.md after my changes”claude-md-updater agent
”Audit my existing project’s structure”workspace-advisor agent
Any “How does X work in [large codebase]?“built-in Explore (Task tool)
Pre-implementation design for multi-file changesbuilt-in Plan (Task tool)
Complex task the user expects will need multiple stepsgeneral-purpose (Task tool)

When recommending, use AskUserQuestion with wording like:

“This looks like it needs a [agent] agent — it’ll give us [benefit]. Want me to spawn one?”


README.md for the quick overview, then 01-kernel/PHILOSOPHY.md for the why.

seacow-scaffolder agent (ask first). Or the deterministic scripts/setup.sh if the template is obvious.

skill-writer agent (ask first).

”What’s the difference between skills and agents?”

Section titled “”What’s the difference between skills and agents?””
  • Skill = passive expertise (loaded into same context, no tool use)
  • Agent = active executor (fresh context window, can use tools, returns summary)

Deep dive: 01-kernel/principles/03-skills-vs-agents.md.

An analytical lens (not a folder structure): S-ystem, E-ntity, A-ctivities (Capture/Output/Work), r-elation. It’s a set of questions for thinking about organization. Load seacow-conventions skill for the full framework.

”Can I use this with tools other than Claude Code?”

Section titled “”Can I use this with tools other than Claude Code?””

Yes — AGENTS.md is portable to 20+ AI tools. Key constraint differences:

  • Claude Code: agents cannot spawn other agents (sequential composition only)
  • OpenCode: agents CAN spawn child agents (recursive composition)

Deep dive: 02-stack/decisions/tool-comparison.md or 01-kernel/ARCHITECTURE.md.

”What’s the stratum: field in frontmatter?”

Section titled “”What’s the stratum: field in frontmatter?””

A classification (1–5) indicating how portable a file is:

  • 1 philosophy — always true; anyone can adopt
  • 2 pattern — same shape, different content; named patterns
  • 3 parametric — fill-in-the-blank templates
  • 4 deterministic — drop-in scripts, fixed content
  • 5 instance — specific to this user; not a template

Deep dive: 01-kernel/principles/07-five-strata.md.


Pattern 1: Start a new project from the scaffold

Section titled “Pattern 1: Start a new project from the scaffold”
  1. Decide: deterministic (use scripts/setup.sh) or interview (use seacow-scaffolder agent).
  2. If interview → ask first, then spawn agent.
  3. Scaffolder creates structure based on user’s context.
  1. Identify the domain/task type.
  2. Ask: “Want me to spawn skill-writer for [domain]?”
  3. If yes → agent creates skill in correct format.
  4. Verify the description covers the right activation contexts.
  1. Identify the repeatable work.
  2. Ask: “Want me to spawn agent-writer for this?”
  3. If yes → agent creates file with appropriate tools and skills.
  1. Load seacow-conventions skill.
  2. Ask SEACOW questions about the situation.
  3. Design YOUR structure based on YOUR answers — don’t copy, think.

Pattern 5: Scaffold maintenance after changes

Section titled “Pattern 5: Scaffold maintenance after changes”
  1. User adds/modifies a skill/agent.
  2. Remind: “Remember to update CLAUDE.md / AGENTS.md.”
  3. Offer: “Want me to spawn claude-md-updater or workflow-expert to check docs?”

After user adds an agent:

  • “Remember to add it to AGENTS.md and CLAUDE.md’s Available Agents section.”
  • Or: “Want me to spawn workflow-expert to update docs?”

After user adds a skill:

  • “Skills should be at 01-kernel/skills/<name>/SKILL.md with YAML frontmatter (name, description, stratum). Remember to update CLAUDE.md’s skills list.”

After structural changes:

  • “Want me to spawn claude-md-updater to check if docs need updating?”

After a research insight worth keeping:

  • “Is this methodology (fits in a tier) or knowledge work (exits to external vault)? If methodology, which tier — kernel (universal), stack (your toolkit), or work (your specific patterns)?”

Every markdown file in a tier should have:

---
title: <page title>
description: <one-line purpose>
stratum: <1-5>
status: <research | planning | stable | parked>
tags: []
date: YYYY-MM-DD
---

Obsidian-style [[path/to/file]] works via remark-wiki-link. Prefer wikilinks for cross-references; regular markdown links for external URLs.

Each section directory can have an index.md for sidebar navigation and orientation.

  • Skills: 01-kernel/skills/<name>/SKILL.md
  • Agents: 01-kernel/agents/<name>.md
  • Commands: 01-kernel/commands/<name>.md
  • Principles: 01-kernel/principles/NN-<name>.md (numeric prefix for sidebar order)
  • Tier zone folders: NN-<name>/ (dash separator, matches 00-inbox/ existing convention)

┌─────────────────────────────────────────────────────────────────┐
│ SHOULD I OFFLOAD THIS TO A SUBAGENT? │
├─────────────────────────────────────────────────────────────────┤
│ Will this require reading more than 3-5 files? │
│ ├── YES → Spawn subagent (Explore/Plan/specialist) │
│ └── NO → Do it directly │
│ │
│ Am I asking "how does X work in this codebase?" │
│ ├── YES → Spawn Explore agent │
│ └── NO → Continue │
│ │
│ Is this a multi-step implementation with architectural choice? │
│ ├── YES → Spawn Plan agent first, review, then implement │
│ └── NO → Do it directly │
└─────────────────────────────────────────────────────────────────┘

Agents spawn with fresh context. When the main context is approaching ~40% used on exploration, offloading to an agent prevents further bloat. The agent returns a compressed summary; main context stays clean for synthesis.

Main → Agent A (explores /docs/) → returns summary
→ Agent B (explores /src/) → returns summary
→ Main synthesizes with clean context

Each level distills. Parents receive only summaries.

Deep dive on delegation decisions: delegation-advisor skill (also works with ask-first pattern).


  • Check path: 01-kernel/agents/<name>.md (or the .claude/agents/ symlink)
  • Verify YAML frontmatter present
  • Ensure description field exists (Claude uses this for auto-invocation)
  • Skills need subdirectory structure: 01-kernel/skills/<name>/SKILL.md
  • SKILL.md must have YAML frontmatter (name, description, stratum)
  • Claude uses description for semantic matching (NOT keyword matching)
  • Review skill’s description — does it cover the contexts you’re activating in?

”Structure feels wrong for my context”

Section titled “”Structure feels wrong for my context””
  • Examples are not prescriptions
  • Use SEACOW thinking to design YOUR structure
  • The tier structure (kernel/stack/work) is THIS repo’s decomposition — your project may need a different decomposition

Start with:

  1. README.md (5 min)
  2. 01-kernel/PHILOSOPHY.md (10 min)
  3. Try seacow-scaffolder on a small project

Learn by doing, not reading everything first.


ComponentRole
workflow-scaffold (this skill)Concierge — answers directly or ASKS to delegate
workflow-expert (agent)Specialist — deep investigation of scaffold content
seacow-scaffolder (agent)Builder — creates NEW structures elsewhere
delegation-advisor (skill)Focused companion — when the delegation decision itself is the question

This skill is the lightweight trigger. The agents do the heavy lifting.


  • 01-kernel/PHILOSOPHY.md — invariants of the system
  • 01-kernel/principles/ — 10 principle pages (philosophy)
  • 01-kernel/patterns/ — named patterns (SEACOW lens, meta-agent pattern, etc.)
  • 01-kernel/ARCHITECTURE.md — tool composability details
  • seacow-conventions skill — SEACOW framework deep dive
  • skill-patterns / agent-patterns skills — designing new components
  • delegation-advisor skill — when delegation is the right call

Merged from workflow-guide + workflow-meta during Phase 1 restructure (2026-04-17). The two previous skills had near-identical purposes (navigation + meta-question handling + ask-first delegation); combining into one resolved the redundancy.