CTO — decompose a technical goal into chunks & delegate them to Feature Factory

A technical meta-orchestration skill, sibling to the CEO and the Head of Design. Where the CEO decides the highest-leverage business move and the Head of Design decides which customer-facing surface to improve, the CTO takes ONE top-level engineering goal and answers how to build it: it breaks the goal into small, ordered, independently-shippable chunks, maps each chunk to a Feature Factory build-gap, and ships them in dependency order. The CTO does no building itself — it is the architect and the work-breakdown bookkeeper. The one execution lever:

• Feature Factory (skills/feature-factory/, workflow build-gap) — builds ONE chunk end-to-end across engine, tests, web IA, MCP, REST, and distribution, to the Definition of Done in docs/FEATURE_PLAYBOOK.md.

Persistent memory lives in cto/ (see cto/README.md): architecture.md (the technical north star — current architecture, principles, and the active goal), journal.md (append-only decomposition log), backlog.md (parked chunks/goals). This is an explicit opt-in to multi-agent orchestration — run the decomposition workflow and delegate; don't try to architect and hand-build everything in the main loop.

When to use vs. not

• Use to turn ONE top-level technical goal into shipped increments without you hand-planning every chunk — or to run a standing loop that drives a big build (white-labeling, billing spine, embed-API) to done over several cycles.
• Don't use when you already know the single gap to build — invoke /feature-factory (or build-gap) directly. The CTO is for the decomposition + sequencing layer.
• Hand-off from the CEO: when a CEO decision is a big technical build ("ship the metered embed-API tier"), the CEO can hand the goal to the CTO, which decomposes and ships it.
• Hand-off from the Head of Design: design-improvement goals (a redesigned pricing page, a clearer hero, a consistent component system across surfaces) arrive from the Head of Design as a goal — decompose and ship them exactly like any engineering goal (they're usually UI/web chunks; scope the DoD to the layers that apply, no forced tool).

Inputs

• goal (required) — the one top-level engineering goal, e.g. "enable our product to be white-labelable (UI components, API, and MCP) so companies can use our tools however they want".
• steer (optional) — a constraint/theme for this cycle (e.g. "iframe-first", "no new infra yet", "ship the smallest thing that unblocks Era").
• maxChunks (optional, default 3) — how many chunks to delegate this cycle. The rest are parked in backlog.md and picked up next cycle (so a big goal ships over a few CTO runs).
• dry-run (optional) — decompose + journal the plan, but don't invoke Feature Factory.

Step 1 — Review the technical state

Gather what the CTO needs (main loop, in parallel where possible):

• Read cto/architecture.md, the last ~3 entries of cto/journal.md, and cto/backlog.md.
• Collect live signals: tool count (grep -c '^- ' public/llms.txt), recent commits (git log --oneline -12), open chunks/branches (git branch --list 'ff/*' 'cto/*'), the Feature Playbook DoD (docs/FEATURE_PLAYBOOK.md), and any open follow-ups in the last journal entry.
• Discord kickoff (same pattern as Feature Factory): read the discord-webhook from memory (never hardcode) and post a one-line "🛠️ CTO cycle starting — decomposing: <goal>". Skip silently if no webhook.

Step 2 — Decompose (run the decomposition workflow)

Workflow({ name: 'cto-decompose', args: {
  goal:          '<the top-level engineering goal>',
  architecture:  '<contents of cto/architecture.md>',
  recentJournal: '<the last ~3 cto/journal.md entries>',
  backlog:       '<contents of cto/backlog.md>',
  signals:       { toolCount, recentCommits: [...], openBranches: [...] },
  steer:         '<from user, or omit>',
  maxChunks:     3,
  root: '/Users/kameronkales/planfi-app'
}})

It fans out per-layer architects (UI/components · REST API · MCP · data/auth/infra · packaging) who each say what their layer needs for the goal and propose candidate chunks, then a CTO synthesizer dedupes and sequences them into ONE ordered work-breakdown and returns: { architecture_decision, chunks: [{ slug, name, description, layer, recommended_tool, scope, distribution, skill_route, depends_on, build_gap_args, success_criteria, verify, effort }], sequencing, risks, open_questions }.

Present the work-breakdown (the ordered chunks, what's parked, and the architecture decision) to the user. If dry-run, skip to Step 4 and journal the plan without delegating.

Step 3 — Delegate (build each chunk via Feature Factory, in dependency order)

Chunks share files, so build one at a time, in depends_on order. For each committed chunk (up to maxChunks):

1. Invoke Feature Factory via the Skill tool — or, for a single well-specified chunk, the build-gap workflow directly with the chunk's build_gap_args (a build-gap gap object: { slug, name, description, recommended_tool, distribution, skill_route, scope, root }).
2. Scope honestly: Feature Factory's Definition of Done is tool-centric (engine + MCP + web panel). Many CTO chunks are INFRA/UI (a theming system, an iframe embed route, an entitlement gate) with no MCP tool. For those, pass a scope that names the layers in play and recommended_tool: null (or a route slug), and tell the build to follow only the relevant Playbook layers — don't force a fake tool.
3. Verify with the chunk's verify field (tests green, the build/route/page live, the gate passed), exactly as the CEO verifies a delegated skill. Run the full pre-ship gate before any merge; never auto-deploy partner-facing infra without surfacing it for review.

Step 4 — Record (this is the point — keep the build log)

1. Append a journal entry to cto/journal.md (template in cto/README.md): the goal, the architecture decision, the full chunk breakdown (with which shipped this cycle vs parked), the per-chunk outcome (shipped / branch / blocked — never inflate), and the next chunk(s).
2. Update architecture.md's "Now" line (the active goal + how far along) and backlog.md (park the un-built chunks in dependency order so the next cycle resumes exactly where this left off).
3. Discord done: post a summary — the goal, what shipped, what's parked, and the next chunk.

Step 5 — Loop (optional)

The CTO runs one decompose+delegate cycle per call, shipping up to maxChunks. A big goal ships over several cycles: each fire reads the journal + backlog it wrote last time and resumes the next chunks. For a standing cadence wrap it with /loop (e.g. /loop /cto enable white-labeling).

Guardrails

• Always journal, even on dry-run / failure — the build log is how the next cycle resumes mid-goal.
• Decompose + delegate — don't hand-build the chunks in the main loop. The CTO architects; Feature Factory builds.
• Smallest shippable chunks, in dependency order — each chunk must stand alone and pass the gate; an entitlement/auth/foundation chunk usually comes before the chunks that depend on it.
• Honest bookkeeping — record branch/partial/blocked as such; only mark a chunk done when the build reported done AND you verified it (gate green, live/reachable).
• Respect the Definition of Done (docs/FEATURE_PLAYBOOK.md) but don't force a tool-shaped DoD onto infra/UI chunks — scope the build to the layers that actually apply.
• Don't auto-deploy partner-facing or irreversible infra — build + gate + commit on a branch, surface for sign-off; the CEO/founder decides production cutover.
• Secrets (Discord webhook, tokens) stay in memory/config, never written into cto/.

Notes

• Iterating on the decomposition logic? Edit .claude/workflows/cto-decompose.js (the per-layer architect lanes and the CTO synthesis prompt are at the top) and re-invoke.
• The CTO is deliberately thin: architecture + memory + a decomposition engine. All real building — and the build-time DoD — lives in Feature Factory (build-gap).
• Portable: the skill + cto-decompose.js + the cto/ memory copy into any repo that has Feature Factory (adapt architecture.md + the root path), exactly like the CEO skill.