Superpowers + Compound V
The two-plugin toolchain for shipping real work with an AI squad: brainstorm a spec, audit the ground, plan a disjoint partition, then route every task to the right model and backend in parallel behind hard review gates, and it remembers what it learns.
A The golden path
idea → merged
One-time per project: /v:init, detect backends, pick a routing stance, save config. Then, per feature:
Brainstorm the spec superpowers
Say brainstorm <idea> (or "let's build X"). One-question dialogue + 2-3 approaches → an approved spec in docs/superpowers/specs/. Hard gate: no code before the design is approved.
Pre-flight audits the real "onboard"
Three agents in parallel case the territory, code-archaeologist (what exists), domain-expert (the rules), doc-validator (library/API reality). /v:archaeology runs just the code audit.
Plan a partition superpowers
Spec → a bite-sized plan with a verified Partition Map: disjoint write-paths per task, so agents never collide.
Cross-model plan review by stakes
/v:review-plan, a different model family (Codex) adversarially reviews the plan read-only; you arbitrate. Run it for security / auth / payments / migrations, big partitions, or architectural change.
Materialize the manifest
/v:orchestrate writes manifest.yaml + state.json. It does not dispatch, inspect or edit first.
Dispatch in parallel
/v:dispatch, batched multi-backend execution. partition-reviewer + validate-manifest gate first; the git-diff scope gate runs after every job. A bare plan path auto-materializes.
Recover & inspect
/v:resume reconciles state against git (git wins) and re-runs only incomplete jobs. /v:status shows the phase, a per-job table, and backend health.
Review gate → rework → merge
Three passes, SPEC → QUALITY → INTEGRATION, checked by a different family. Any issue bounces the job back to the bench for a fix, then re-review. Only all-green merges. No "close enough".
Phases: SPEC_READY → PREFLIGHT_DONE → PARTITION_VERIFIED → DISPATCHED → COLLECTED → REVIEWED → MERGED · /v:epic chains several features into one autonomous, resumable, dependency-ordered build, each runs the full path above.
B Compound V commands
/v:*
.claude/compound-v.json.docs/superpowers into the local V-memory cache.C Superpowers skills
not commands · auto-fire
| Skill | Trigger & purpose |
|---|---|
| brainstorming | "let's build X" → dialogue → approved spec. No code before approval. |
| writing-plans | spec → task-by-task plan (paths, code, TDD, commits). |
| subagent-driven-development | recommended fresh subagent per task + two-stage review. |
| executing-plans | sequential execution with checkpoints (separate session). |
| test-driven-development | RED → GREEN → REFACTOR. No prod code without a failing test. |
| verification-before-completion | run the real command now before claiming done. |
| systematic-debugging | root cause before any fix; 3+ failed fixes = architecture problem. |
| requesting / receiving-code-review | reviewer subagent on a real diff; verify feedback before applying. |
| using-git-worktrees | isolated workspace; detect existing isolation first. |
| finishing-a-development-branch | merge / PR / keep / discard, then clean up. |
D V-memory
semantic project recall
A local-first recall layer over docs/superpowers/** prose, specs, plans, reviews, audits, lessons. It answers "have we seen this before?" by meaning, not filename grep, and accumulates as the project grows: every plan and postmortem you commit becomes searchable on the next refresh. It stays subordinate, evidence for planning & review, never a routing input.
FTS5, BM25, always on, pure stdlib, offline, instant. The dependable substrate.
Embeddings, multilingual-e5-small, opt-in, cached outside the repo. Kicks in for large corpora; absent → silently FTS5-only.
recall-check: a file pattern with ≥2 prior blocked/error jobs → verdict tighten (force worktree / +1 review / fold into Task 0). Conservative-only, it can only tighten.
# recall past context as evidence /v:remember "notion oauth token refresh" # reindex (+ optional semantic lane) /v:memory-refresh --with-embeddings # the engine, directly python3 scripts/compound-v-memory.py bootstrap # the ONLY network step python3 scripts/compound-v-memory.py refresh --with-embeddings python3 scripts/compound-v-memory.py search "<q>" --intent planning python3 scripts/compound-v-memory.py doctor # index / venv / health
E Backends & routing
who does what
V routes by tier, not a hardcoded model, the concrete model resolves at dispatch from a refreshable map (/v:models), so it survives model churn. haiku is never used.
mechanical single-file edits, docs, i18n, well-specified CRUD.
bounded feature build, including large-isolated Codex work.
architecture, sensitive surfaces, new tests, external APIs, every reviewer, Task 0.
| Backend | Trust | Role |
|---|---|---|
| claude | kernel-trusted, in-process | default; Task 0, sensitive surfaces, reviewers, docs |
| codex | kernel-sandboxed worktree | recommended ally for large-isolated build |
| antigravity | lower-trust worktree | opt-in extra hands, trusted surface only |
| cursor | lower-trust worktree | same tier & caveats as antigravity |
- reviewers ⇒ deep (opus)
- codex / antigravity / cursor ⇒ worktree
- run: parallel ⇒ isolation: worktree
- unclear scope ⇒ back to planning
- scope gate: any write outside
write_allowed⇒ blocked, never merged
Stances (pick at /v:init): Balanced (default, with Codex) · Claude-only (Codex absent → rows collapse to claude/deep) · Conservative · Cost-aware, but sensitive surfaces & reviewers stay deep in every stance.
F Optimal setup
subscriptions
Funds the deep / reviewer / Task-0 work, architecture, every sensitive surface, all reviewers. This is what's never cost-optimized in any stance, so it needs the highest-trust, in-process backend.
Runs as its own sandboxed process/worktree, so large builds parallelize off Claude's plate, and a different model family gives you the error-decorrelation the review layer is built on. Cheaper than a second Opus seat.
Extra parallel capacity, worth adding only once you have trusted, narrow-blast-radius build work, no kernel write-confinement, so the gate detects leaks but can't prevent out-of-worktree side effects.
The plugin forbids fabricated cost/token numbers ("anti-ruflo"), so this stays at the subscription-tier level, the $200 seat funds work that can't be cheapened; the $100 seat funds parallel throughput plus a second opinion.
G Skills for the Universal Creator
founder brain
Compound V is the build engine. These skill packs add the solo founder's brain, product, marketing, thinking and anti-slop copy, so one person can own the whole product.
Marketing Skills →
60+ marketing skills: CRO, copy, SEO, ads, launches, pricing. The whole go-to-market a solo builder owns.
/plugin marketplace add coreyhaines31/marketingskills
Product Manager Skills →
55 PM skills: discovery, RICE/ICE/Kano prioritization, PRDs, roadmaps. Structure for what to build.
/plugin marketplace add deanpeters/Product-Manager-Skills
Thinking Skills →
39 mental models: first principles, inversion, pre-mortem, JTBD, red-team. To pressure-test the concept.
/plugin marketplace add tjboudreaux/cc-thinking-skills
Stop Slop →
Detects and strips AI writing patterns and scores directness. Keeps your PRD and launch copy human.
git clone https://github.com/hardikpandya/stop-slop.git
Brainstorm the spec → audit the ground → plan a disjoint partition → let V route each task to the right model + backend in parallel → gate everything (scope + 3-pass cross-model review) → rework what fails → merge. V-memory quietly remembers all of it, so the next run starts smarter.