j-stack
A Claude Code stack optimized for enterprise PoC delivery.
Higher-fidelity demos in less time.
Defensive deliverables that preempt "did you try X" objections.
Cross-tool memory — resume in Codex / Cursor / ChatGPT when Anthropic limits hit.
Anthropic's AI coding CLI. Not autocomplete — an autonomous agent that reasons about architecture, writes and edits files, runs tests, browses docs, and dispatches parallel subagents. Think of it as a senior engineer who lives in your terminal.
A plugin framework that installs structured workflow skills into Claude Code. Skills enforce how Claude approaches problems: brainstorm before planning, spec before building, test before shipping. j-stack uses brainstorming, writing-plans, TDD, and subagent-driven-development.
A community skill pack with 40+ specialized tools: security audits, CEO-lens scoping, design QA, document generation. j-stack cherry-picks 11 skills that complement Superpowers and injects explicit model directives so expensive models aren't wasted on mechanical work.
j-stack is the opinionated assembly. Raw Superpowers + gstack is powerful but sprawling — 40+ skills with no lane ownership, no model discipline, and no cross-tool memory. j-stack curates the set, routes models by cognitive demand, adds the four custom skills that complete the pipeline, and wires the whole thing into a repeatable workflow.
An opinionated assembly of three tools — curated, model-routed, and extended with four custom skills that don't exist anywhere else.
j-stack assembles Claude Code + Superpowers + cherry-picked gstack skills + four custom skills into an opinionated workflow pipeline. Each phase is owned by a specific skill — nothing overlaps, nothing conflicts. The result is a repeatable system for shipping polished, defensible PoC artifacts.
Writing-plans locks a full implementation spec before any code runs. Subagents implement against the spec, not a vague prompt.
The TDD skill enforces red-green-refactor discipline. Tests are written first, against the spec, then implementation follows.
OpenAI Codex independently reviews Claude's work. Convergent findings from two AI vendors are the strongest signal you can put in front of a skeptical architect.
A .planning/ wiki keeps every decision readable by Claude Code, Codex, Cursor, ChatGPT, and Gemini. Hitting usage limits mid-engagement is a handoff, not a loss.
EXPAND + REFINE exist before a line of spec is written. The pipeline cannot reach PLAN without completing both.
prior-art-survey dispatches three parallel scouts — OSS, libraries, patterns — and produces a named, reasoned landscape. Not a gut call.
Each stage must complete before the next begins. Nothing in BUILD runs without a locked PLAN output.
/cso files an OWASP/STRIDE finding set. second-opinion adds a cross-vendor risk matrix. Both are artifacts, not chat notes.
/document-release generates docs from the diff. handoff-snapshot captures state. No documentation sprint at the end.
subagent-driven-development and prior-art-survey's scouts both parallelize work the way a PM schedules concurrent engineering tracks.
Nothing in BUILD starts until PLAN is locked. The pipeline enforces four phases before a line of implementation code is written: reframe the problem → pressure-test the approach → survey the prior art → spec the implementation.
writing-plans produces a full implementation plan reviewed by Opus. Subagents in BUILD implement against that spec — not a vague prompt, not a wish.
The superpowers:test-driven-development skill enforces red-green-refactor inside the build phase. Tests are written first against the spec; implementation follows only once the test is failing for the right reason.
subagent-driven-development parallelizes implementation across multiple Claude instances, each operating against the same locked spec and test suite.
Vision doc traces the founder-lens reframe and every major decision made during scoping.
Parallel OSS, library, and pattern scouts produce a named, reasoned landscape. "Did you try X" is answered before it's asked.
OWASP Top 10 + STRIDE security review runs before the demo. Findings are filed by severity with recommended mitigations.
QA audits the running app against the spec. The TDD test suite from BUILD provides the passing-green baseline.
OpenAI Codex independently reviews Claude's work. A convergence/divergence matrix is the deliverable — two AI vendors agreeing is a stronger signal than one.
All five answers assembled into one executive-ready document before the demo. The meeting shifts from "defend the work" to "discuss what's next."
SP = Superpowers
Never lose context at usage limits.
Each PoC project gets a .planning/ wiki with schema files for every major tool — CLAUDE.md, AGENTS.md, .cursor/rules, chatgpt-brief.md. When Anthropic limits hit, run handoff-snapshot and paste the continuation prompt into your next tool.