SPECTRE

The second major feature group is session memory and context management, provided by the /spectre:handoff and /spectre:forget commands. Session memory allows SPECTRE to maintain and accumulate context across multiple Claude Code sessions. When the context window approaches 160k tokens, users run /spectre:handoff to generate a status report that automatically loads into the next session. The /spectre:forget command clears memory when switching to a completely different topic. This prevents the agent from losing track of progress, decisions, and pending tasks. Subagents like @spectre:sync review the last three status reports and merge them into a continuous memory. This ensures that complex multi-session feature development remains coherent, and users can pick up exactly where they left off without re-explaining the entire context each time.

The third feature group is the Evaluate and learning system, which combines architecture review with knowledge capture. The /spectre:evaluate command triggers an Opus 4.6 subagent to perform a principal-level architecture review of completed work while simultaneously capturing durable project knowledge into reusable skills. These skills auto-load in future sessions through a SessionStart hook that injects the project's knowledge registry into context. When a task matches a trigger word (e.g., "auth" triggers a feature-auth-flows skill), Claude loads the full skill before searching the codebase. Captured knowledge includes gotchas, architectural decisions, feature dossiers, reusable patterns, and multi-step procedures. This knowledge compounds across sessions, making each subsequent interaction faster and more accurate. Users can also run independent commands like /spectre:learn, /spectre:architecture_review, or /spectre:recall to access specific knowledge on demand.

SPECTRE works through an orchestrated workflow that begins with a single slash command and progresses through each phase with automated suggestions. Users typically start with /spectre:scope to define requirements, constraints, and success criteria. The agent asks targeted questions—only those it cannot answer from codebase research—to produce a crisp scope document. Next, /spectre:plan researches the codebase and creates an implementation plan, which can be a high-level technical design or a concise set of tasks. Execution with /spectre:execute uses parallel subagents (@spectre:dev, @spectre:tester, etc.) to implement features in waves, each subagent producing completion reports for the next. Validation is handled by /spectre:validate, which breaks down tasks and dispatches verification subagents. Throughout, /spectre:handoff keeps context windows clean, and /spectre:ship provides autonomous end-to-end feature delivery from brain dump to pull request with zero confirmation gates.

Concrete use cases include building Subspace, a 250,000-line Tauri/Rust/React desktop application, and New June, a React Native AI Agent with a GPS rangefinder for golfers. Both were developed using SPECTRE as the daily driver workflow. A typical day for the creator involves starting with /spectre:scope for new features, refining UX with /spectre:ux if needed, then planning with /spectre:plan, executing with /spectre:execute, cleaning with /spectre:clean, testing with /spectre:test, rebasing with /spectre:rebase, and finally evaluating with /spectre:evaluate. For low-complexity tasks, /spectre:ship autonomously scopes, implements with TDD, sweeps, rebases, and opens a PR. Bugs are handled with /spectre:fix for structured debugging. This workflow enables an ex-Meta, ex-Amazon technical product manager to build, ship, and iterate on products with 100x the complexity of anything built before, delivering working code quickly and iterating from a solid baseline.

SPECTRE is designed for product builders—technical product managers, indie hackers, full-stack developers, and engineering leads who use Claude Code or Codex as their primary coding agent. It works on both new projects and massive codebases of hundreds of thousands of lines, supporting websites, React Native apps, native desktop apps, and personal software. The tool is installed via marketplace commands within Claude Code or a simple npx command for Codex. It is open source under the MIT license, with an active repository on GitHub where users can report issues and contribute. The core philosophy is that great inputs produce great outputs, and that structured workflows are the easiest way to provide the specificity LLMs need. By reducing ambiguity and accumulating institutional memory, SPECTRE makes high-quality, repeatable AI-driven development accessible to anyone building real products. Whether you are shipping your first feature or managing a complex codebase, SPECTRE helps you ship faster with confidence.