Idea To Build

Idea → Build Pipeline (global, model-tiered)

A cross-surface pipeline that separates ideation, specification, architecture, and implementation by model tier and surface. The durable state between stages is the ticket tracker — never a chat transcript. Any new session (any device) resumes by reading the tracker.

Why tiers: Sonnet is a cheap, fast divergent thinker — ideal brainstorm partner. Fable is the scarce, expensive judgment — spend it ONLY on architecture, planning, review gating, and orchestration. Opus 4.8 agents are strong implementers at a fraction of Fable's cost. Fable never writes the code; it instructs agents who do. Same output quality, fraction of the price.

The four stages

Hard gate: Stages 3–4 never happen in a plain chat. They require the project file — Fable must see the real repo (what already exists, conventions, migrations half-done) alongside the tickets.

Stage 1 — Brainstorm (Sonnet, anywhere)

• Role: thought partner. Ask probing questions, run WebSearch/research on demand, surface trade-offs, converge on scope. Brain-dump friendly.
• In Claude Code: switch with /model sonnet, then use superpowers:brainstorming (or product-management:product-brainstorming for product/feature ideation, per global precedence).
• On claude.ai/phone: plain chat works — no repo needed. Ensure the tracker connector is enabled on claude.ai so Stage 2 can run from the same conversation.
• Done when: the idea's scope, non-goals, and key decisions are settled enough that ticketizing wouldn't invent anything.

Stage 2 — Ticketize (same surface; tracker required)

Convert everything decided into tickets. Every ticket must carry:

1. Dependencies / blockers — what it relies on, linked explicitly (blocked-by relations).
2. Executor tag — agent-ok (an AI agent can complete it end-to-end) or human-in-loop (needs a human decision, credential, purchase, or irreversible action). Be honest; when unsure, tag human-in-loop.
3. Acceptance criteria — verifiable, so a Stage-4 agent knows "done".
4. Enough self-contained context to be executed without the brainstorm transcript.

Tracker resolution order (use the first available):

1. Linear MCP (mcp__linear__* / the claude.ai Linear connector) — the canonical choice; supports blocked-by relations + labels natively.
2. Jira (Atlassian MCP: createJiraIssue, createIssueLink for blocks/is-blocked-by).
3. Notion (a tickets database with Relation property for dependencies).
4. Claude Code native tasks (TaskCreate + TaskUpdate addBlockedBy) — fine for solo/single-machine projects; not readable from phone chats.
5. mattpocock to-issues (files in docs/agents/) — repo-native fallback; requires the setup-matt-pocock-skills precondition per global CLAUDE.md.

Stage 3 — Plan (Fable, in the project repo)

1. Pull all relevant tickets from the tracker (list them; read every one).
2. Read the actual repo state — Fable must understand what already exists before planning (greenfield vs brownfield changes everything).
3. Produce the build plan: sequencing from the dependency graph, per-ticket dispatch strategy, what runs in parallel, where the human gates are. Use plan mode / superpowers:writing-plans; use gepetto first for deep upstream architectural research when the build is large/greenfield.
4. Fable's architectural judgment overrides Stage-1/2 choices where they conflict — but it must (a) say so explicitly, (b) write the deviation back to the affected tickets (comment or edit), so the tracker stays the single source of truth. User-locked product decisions are never overridden, only technical approach.

Stage 4 — Execute (Fable = EM; Opus 4.8 agents = hands)

• Fable stays the Engineering Manager (global EM mode applies): it NEVER implements; it dispatches, reviews, integrates.
• Every implementation dispatch pins the worker model: Agent(..., model: "opus") — Opus 4.8. Do NOT let workers silently inherit the session model (inheriting = agents run as Fable = quality-identical but several× the cost). Combine with the right specialist subagent_type from ~/.claude/agents/.
• Reviews are maker≠checker: independent reviewer agent (also model:"opus", fresh context/worktree) per meaningful unit. superpowers:subagent-driven-development + dispatching-parallel-agents govern the mechanics; verification-before-completion closes it.
• Work ticket-by-ticket along the dependency graph; parallelize independent tickets. Commit+push per ticket (durable checkpoints — background agents die; commits don't).
• human-in-loop tickets: STOP and ask the user; never auto-complete them.
• Close/update each ticket in the tracker as it lands (status + link to commit/PR). The tracker must end truthful.

Resumability & hygiene

• The tracker is the memory. A dead session, a new machine, or a phone follow-up all resume by reading tickets — never by reconstructing chat history.
• If the session model isn't Fable at Stage 3/4, tell the user to switch (/model claude-fable-5) rather than proceeding with a weaker architect.
• Cost sanity: Sonnet for volume ideation, Fable tokens only for judgment, Opus for production code. If a Stage-4 task is trivial/mechanical, model:"haiku" is permitted for it.