agentsclimarketplace

Agent build sequence

Skill cardinalconseils/claude-starter/skills/agent-build-sequence

Agentic System Build Sequence — 15-stage ordered methodology for building AI agent systems, autonomous agents, multi-agent orchestration, tool-calling loops. Enforces state machine and tool inventory design before architecture decisions, and architecture before schema. Use for AI agent system, autonomous agent, agent system, multi-agent, LLM economics, tool inventory, capability matrix design.From its SKILL.md

Install
npx -y skills add cardinalconseils/claude-starter --skill agent-build-sequence

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 3 stars3 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.8 KB, as published. Nobody here has run it

Agentic System Build Sequence

Overview

Building an AI agent system out of order is the single most common cause of expensive rework: teams pick an architecture before they know what tools the agent calls, or design a database schema before the state machine that governs the agent's lifecycle exists. This skill is a 15-stage ordered methodology that fixes the order, not the content — most stages are dispatch pointers to CKS agents/skills that already do the work well. The value-add is sequencing:

State machine + tool inventory must exist before architecture. Architecture must exist before schema.

Reversing this order is the recurring failure mode: an architecture decision made without a tool inventory has no idea what latency/cost/failure-mode envelope the system must tolerate, and a schema written before architecture bakes in assumptions (e.g., synchronous vs. event-sourced) that get expensively unwound later.

When to Use

  • Building an AI agent system, autonomous agent, or multi-agent orchestration from scratch
  • A feature description mentions "tool-calling loop", "agentic system", "AI agent", or similar
  • project_type: ai-agent-system is set in .kickstart/state.md or .bootstrap/scan-context.md
  • Any point where architecture, schema, or API contract decisions are being made for an agent system and no state machine or tool inventory exists yet

When NOT to Use

  • Simple LLM feature additions to an existing non-agentic app (a single prompt call, a classifier) — this is 15-stage overkill for a one-shot LLM call
  • The project already has a mature state machine, tool inventory, and architecture — re-running the full sequence for a small feature addition inside an existing agent system is unnecessary
  • project_type is anything other than ai-agent-system and no agent-system keywords matched

The 15 Stages

Full gate-check detail (what artifact each stage produces, exact dispatch target, prerequisites) lives in workflows/build-sequence.md. This table is the map — read the workflow file before running any stage.

#StageProducesDispatch Target
1Idea / Problem StatementRefined pitchcks:ideate
2Monetization / Validation.monetize/evaluation.mdmonetize-discoverer → monetize-researcher → monetize-evaluator
3PRDCONTEXT.mdcks:discover / agents/prd-discoverer.md
4State Machine DesignState/transition tablesskills/orchestration/workflows/state-machine.md
5Tool Inventory / Capability Matrixreferences/tool-inventory-template.md filled inNet-new (this skill)
6Architecture DecisionsARCHITECTURE.md / ADRsagents/architecture-generator.md (only after 4+5 exist)
7Memory ArchitectureMemory design noteagents/honcho-integrator.md, skills/user-memory, skills/conversation-state, skills/honcho-memory
8LLM Economics.decisions/LLM-ECONOMICS.mdThin synthesis of skills/openrouter/workflows/model-research.md, skills/luv-model-routing/SKILL.md, agents/cost-analyzer.md / agents/cost-researcher.md
9ObservabilityTelemetry + eval wiring.claude/rules/telemetry.md, .claude/rules/harness-evals.md, cks:observe
10Error Handling / RecoveryDLQ/retry/circuit-breaker ADRs.claude/rules/arch-patterns.md
11System DesignScreens / component specs (if any UI)agents/prd-designer.md
12API Contract DesignAPI.md (MCP tool defs)skills/kickstart/workflows/design.md Contract Format table
13ERDERD.mdagents/db-erd.md
14Schema ValidationSchema reviewagents/db-investigator.md / skills/database-design/SKILL.md
15Implementation PlanPLAN.mdagents/prd-planner.md

Ordering Constraint (the actual value-add)

Stages 4 and 5 (state machine, tool inventory) MUST exist before Stage 6 (architecture) proceeds. Stage 6 must exist before Stage 14 (schema validation). This is enforced as a Gap Check — a prose checklist run by the orchestrating agent before declaring any stage complete, not a scripted hook. See workflows/build-sequence.md § Gap Check. Per .claude/rules/setup-philosophy.md's bucket test, detecting "was this decided before its prerequisite existed" requires reading intent and timeline — not enumerable as a regex, so it stays judgment-guided.

Net-New Content

Only two pieces of this skill are genuinely new (everything else is a dispatch pointer):

  • Stage 5 — Tool Inventory template (references/tool-inventory-template.md): no existing CKS artifact captures per-tool latency/cost/failure-mode/fallback. This is the missing piece the framework actually needed.
  • Stage 8 — LLM Economics synthesis: not a new agent, just a thin doc that pulls together three already-existing cost/routing skills into one .decisions/LLM-ECONOMICS.md output.

Common Rationalizations

RationalizationReality
"We know the architecture already, let's skip the tool inventory"The tool inventory is what tells you whether that architecture can survive a 3s-p95 tool call or a $0.02/call cost. Skipping it means guessing.
"State machine design is overkill for a simple agent loop"Even a simple agent has states: idle, tool-calling, waiting, done, failed. Naming them now is cheaper than debugging an undefined state later.
"This is just a prompt change, not a new agent system"Then this skill doesn't apply — see When NOT to Use. Don't force 15 stages onto a one-shot LLM call.
"Architecture-generator can figure out the tool inventory itself"It can't — Mode 3 of architecture-generator.md takes tool inventory and state machine as inputs, it doesn't derive them.
"The Gap Check should be a hook so it can't be skipped"Detecting a stage-ordering violation requires reading intent and history, not a regex match. This is deliberately guided, not scripted — see .claude/rules/setup-philosophy.md.

Verification

  • Every stage that was run produced its named artifact at the expected path
  • Stage 4 (state machine) and Stage 5 (tool inventory) artifacts exist before Stage 6 (architecture) started
  • Stage 6 (architecture) artifact exists before Stage 14 (schema validation) started
  • Gap Check (workflow file) was run before declaring the sequence complete
  • Any stage skipped as "N/A" has a one-line reason recorded, not silently omitted

Keep looking

Skills are one crate of 609,856. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.