System Architecture — Runtime-Independent Design
Clearfork C-Suite is an AI-powered executive performance system for CEOs and founders of mid-market companies ($5M–$100M). It consists of a generic agent control plane — Mission Control — with a domain-specific overlay — C-Suite — installed on top.
The CEO experiences a strategic advisor and performance coach. Behind the scenes, the system maps their business, tracks commitments, logs automation opportunities, and builds a compounding intelligence layer that gets more valuable over time.
The system is not one giant prompt. It is a layered architecture where identity, operating rules, workflows, and state are separated cleanly so each layer can evolve independently.
A third layer — the Client Instance — holds per-client data: CEO profile, people map, process map, tech stack, priorities, commitments, opportunities, artifacts, and configuration.
The coach has two simultaneous jobs in every conversation:
Be the CEO's thinking partner. Draw out what they already know. Ask the questions nobody else asks. Be the one person who's always caught up.
Map the business through conversation. When automation or AI opportunities surface, log them — whether the CEO engages with them or not.
Start by drawing out what the CEO already knows. When they need strategic thinking, bring the best available and frame it in their world.
Build the full picture of the business before going deep on any one problem. Keep going until they run dry. Only then prioritize where to dig in.
The coach never architects workflows, drafts technical specs, or shifts into developer mode during a coaching conversation. Log the opportunity, stay in coaching.
Don't get cute redirecting to Socratic questions when they want a direct answer. Be the most valuable strategic mind in the room.
Don't stall by redirecting to coaching. Ask for context if needed, execute with radical specificity, then use what the work reveals as a coaching door if the moment is right.
Who is the coach? Mission, values, voice, relationship stance, coaching method, characteristic moves, hard human boundaries. Does NOT contain logging instructions, schemas, or post-session mechanics.
The hidden runtime rules. Mission hierarchy, behavioral rules, escalation boundaries, delegation rules, transformation mode behavior, honesty rules. This is the operating law. It is not the soul.
What the system knows about this CEO and their business. CEO Profile, Relationship Signals, People Map, Process Map, Tech Stack, Data Landscape, Priorities, Commitments, Friction Log, Opportunity Register, Artifacts.
Repeatable bounded operations: session closeout, opportunity scoring, commitment review, company model refresh, artifact indexing, delegate to builder, delegate to research. Not personality. Not identity.
Coach, Research, Builder. The coach does not do everything. The coach may call the other agents. The coach does not become them. This separation is a load-bearing wall.
Load relevant client state — recent commitments, active priorities, relevant people/process/tech/data slices, relationship signals, transformation mode. Only what's relevant. Not the full database.
Governed by identity (SOUL) + operating contract (AGENTS) + available context. Trusted operator. Highly specific, emotionally intelligent, commercially sharp, never salesy.
The CEO sees Step 2. The system runs Steps 3–6.
Stay coaching. If in doubt, keep coaching.
Stop coaching. Do not diagnose, minimize, or prescribe. Flag to Clearfork immediately. "Can't sleep because of the Danny decision" = coaching. "I haven't felt anything in weeks" = clinical.
Give your best strategic thinking. Be the most useful mind in the room. Caveat appropriately. Never refuse to engage just because a professional should eventually be involved.
Go quiet. Don't argue, defend, or plead. Flag to Clearfork for human re-engagement.
OpenClaw provides out of the box: session management, memory (SQLite + embeddings), prompt caching (90% cost savings), bootstrap file system with priority-ordered loading, skills framework, MCP tool integration, heartbeat/proactive outreach, context compaction, and a working UI.
Building on Agent SDK or Hermes would require rebuilding all of this infrastructure before a single coaching conversation could happen. OpenClaw is the fastest path to a working product. Treat it as substrate, not product logic.
Mission, values, voice, coaching method, characteristic moves. Does not contain behavioral rules, logging instructions, schemas, or post-session mechanics.
Mission hierarchy, behavioral rules, opportunity handling, escalation boundaries, delegation rules, honesty rules. Sub-agents receive AGENTS.md + TOOLS.md only — the right separation for free.
MVP: session-closeout, artifact-index. Phase 2: opportunity-score, commitment-review, company-model-refresh, delegate-builder, delegate-research.
MVP: coach agent only. Phase 2: research and builder sub-agents. Sub-agents inherit AGENTS.md + TOOLS.md — mission hierarchy and honesty rules propagate automatically.
| Priority | File | Purpose | Cache |
|---|---|---|---|
| 1 | AGENTS.md | Operating contract, hidden rules | ✓ Cached |
| 2 | SOUL.md | Identity, voice, values | ✓ Cached |
| 3 | TOOLS.md | Tool usage guidance | ✓ Cached |
| 4 | IDENTITY.md | Additional identity (unused for now) | ✓ Cached |
| 5 | USER.md | Client profile — loaded every turn | ✓ Cached |
| 6 | HEARTBEAT.md | Proactive outreach rules | ✓ Cached |
| 7 | BOOTSTRAP.md | Additional context (unused for now) | ✓ Cached |
| 8 | MEMORY.md | Memory index / pointers | ✓ Cached |
All bootstrap files are cached. All must be byte-identical across turns for cache hits. No dynamic content in any of them. Combined budget target: under 60,000 chars (~15,000 tokens).
OpenClaw loads bootstrap files automatically. Memory system retrieves relevant state files on demand based on conversation content.
Governed by SOUL.md + AGENTS.md + loaded context.
OpenClaw produces one response. Workaround: inline updates + session-closeout skill.
State lives in workspace files. Workaround: structured markdown files with consistent schemas.
No automated scoring pipeline. Workaround: coach logs opportunities with manual assessment fields.
No runtime config switch. Workaround: instruction in AGENTS.md that coach checks USER.md for a transformation_mode field.
None of these block shipping. All of them get solved in Phase 2.
None of these are urgent. All of them are real possibilities as the product grows.
Full control over agent loop, native multi-agent with handoffs, hooks for hidden extraction, MCP, structured outputs, checkpointing. Estimated 150–300 hrs to reach OpenClaw feature parity. Best for: owning the full Mission Control stack.
Better suited to long-running stateful agents with structured memory. Natively understands persistent identity + evolving memory. Less suited if heavy execution/builder capabilities are needed. Similar effort to Agent SDK.
Valid if the gaps don't hurt you. The discipline of keeping product logic portable means migration is always an option, never an emergency.
SOUL.md, AGENTS.md, skills, and state schemas are plain text. They work on any runtime that can read files.
Structured state objects have defined fields. When you move to a database, these become table schemas.
Each skill is a bounded workflow with defined inputs and outputs. They port to any framework that supports procedural prompts.
Don't patch OpenClaw's core prompt assembly. Don't rely on OpenClaw-specific features that have no equivalent elsewhere.
AGENTS.md is where runtime-specific behavior gets encoded. When you change runtimes, AGENTS.md is the primary file that needs updating. SOUL.md should rarely change.