Build features end to end with Spectre's six-phase agentic engineering workflow designed from the ground up to focus your time & effort where it matters the most.
For Claude Code & Codex · Free & MIT
See how it worksSpectre guides you from scope to PR with a crafted six-phase workflow. Every gate, artifact, and handoff is designed to efficiently pass the exact right context where its needed to deliver high quality results.
Spectre is designed specifically so each point of human leverage is maximally utilized. You own the scope, UX, technical design, and final sign off. Agents research, tee up topics, iterate with you until the idea is crisp, then build, validate, and prove. You make the high impact decisions, Agents collaborate.
Every handoff is designed. Iteration gates where you collaborate until aligned. Review gates where you approve to proceed. Adversarial reviews between agents that keep each other honest — no blind trust, just verified output.
Iteration gates where you collaborate until aligned; Review gates where you approve to proceed; meticulously orchestrated subagents workflows in between that provably deliver the scope.
In, out, future, anti-scope — until the edges are crisp. If the experience is fuzzy, define user flows & prototype UI first.
/spectre:scopeIn: inbox, unread, mark-read · Out: email/push (future) · Anti-scope: no prefs UI this phase.
Iteration gate — agent probes, asks, restates iteratively until shape is aligned. Same in UX.
Agent researches the codebase, structures a design, probes and restates with you — then a review gate. You approve the high-level design before tasks exist.
/spectre:planWave 1 T-01 inbox model + API · T-02 team scoping Wave 2 T-03 counts · T-04 mark-read · T-05 inbox UI Wave 3 T-06 batching · T-07 retention
Plan writes goal-prompts.md — paste /goal to start.
Autonomous stretch — subagents build in parallel with only the context they need. Meticulously orchestrated, verified per wave, opposite-runtime review at the end. You don’t babysit.
/spectre:execute · /goalT-03 unread counts ● in progress @spectre:dev T-04 mark-read ○ pending completed T-01 T-02 ✓ verified
Not just passing tests, passing experiences. Your Agent uses the product like a user, and proves it works with screenshots, video, and a visual artfiact for your review.
/spectre:proofNew notification appears · unread matches · mark-read persists · batching coalesces · empty states clear.
screenshots · video · logs · proof.json
Clean up dead code and find duplication, fill risky test gaps, rebase, and open a high-quality PR. The final 5%, one command.
/spectre:ship-itDead code pruned · 4 behavioural tests added · rebased clean.
PR #482 · 9 tasks · 6 criteria · lint clean
Turn what you built into knowledge skills your agents automatically load when relevant. Better context, better results.
/spectre:learnNot notes. A loadable record: the gotcha, decision, or pattern that would otherwise cost debug time next session.
feature-notifications-dossier · gotchas-retention-idempotency patterns-inbox-batching · decisions-postgres-over-sqlite
Search → load by exact ID → update or create → approve → register. Injected at SessionStart, loaded only when subject + use condition match.
Choose your agent and add Spectre in two commands.
/spectre:scope$spectre-scopeUse the command for your agent to scope your first feature.