forge

Implement and VERIFY a feature, fix, or change end-to-end — iterating autonomously until the user's ask is objectively met. Covers UI changes (layout, spacing, typography, colours, copy — verified on a real device via OTA) AND non-UI changes (backend, data-flow, tooling, config — verified via TDD + test suite green + code review). Use this whenever the user asks to build, change, or fix something in any Tessellate app — e.g. "fix the home screen spacing", "add the API endpoint", "the data migration is wrong", "redesign the empty state", "fix the build script". Trigger even when they don't say "feature" or "skill": any change that should be verified before handing back belongs here. For UI changes, the skill measures the result against acceptance criteria from a device screenshot and keeps iterating until it passes. For non-UI changes, acceptance criteria are test-based (Steps 3-4 are skipped; the test suite + code review are the verification bar).

Daily crash and issue monitor for Tessellate apps — triages Sentry crashes and open GitHub Issues, filters dev-mode noise before analysing anything, investigates root cause from stack trace plus source, then auto-fixes and auto-merges only what clears an explicit confidence heuristic. Detects reverts of its own past auto-fixes and applies a cooldown so a bad fix cannot loop. Routes ambiguous fixes and config problems to a human with the decision stated inline. Use whenever the user asks to "check for crashes", "triage Sentry", "what's broken in production", "scan the open issues", "did anything crash overnight", or wants recent errors investigated and fixed. Also fires on passive cues like "is the app healthy" or "anything blow up since the release". Runs daily via the scheduled-tasks MCP; also fires manually. Silence when nothing is wrong is correct output, not a failure.

Spin up a brand-new Tessellate app repo from a short requirements brief — wired to the shared platform from commit one. Use when the user says "new app", "spin up an app", "scaffold a new project", "start a new app called X", or hands over a requirements brief. Creates the repo under Tessellate-Studio, wires the shared libs (rubric-sdk, code-standards, litmus) + the forge plugin, seeds CLAUDE.md from the canonical base + the platform standards + app-specific memory, registers the weekly roadmap-pulse cron, and lands it as ONE initial commit + ONE setup PR (clean history). Seeds essentials only — no pre-built backlog (roadmap-pulse does that later). Always confirms name / package-id / platform before scaffolding; never guesses them.

Three-tiered planning framework for architecture and feature decisions — auto-detects the right tier based on blast radius and reversibility, then runs a structured web-research phase before drafting. ADR for tactical decisions (which library, API pattern, data model — ~5 min). Shape Up Pitch for feature scoping (what to build, what to cut, time appetite — ~30 min). RFD for architecture/design (new system, major refactor, cross-repo change — ~1-2 hours). Planning docs persist at memory/decisions/ as type-prefixed numbered files (adr-001-..., pitch-002-..., rfd-003-...) that become persistent context for future sessions. Use whenever the user asks to "plan", "write an ADR", "pitch this feature", "write an RFD", "should we use X or Y?", "how should we architect this?", "design this system", "what's the right approach for X?", or before building any non-trivial feature. Also auto-triggers from build-feature Step 0 for feature/architecture changes.

Weekly project-management skill that scans planning docs (BACKLOG, RELEASE notes, regression log), runs an honesty pass on stale claims, infers task dependencies, scores open tasks via RICE (Reach × Impact × Confidence / Effort), and produces a prioritized top-5-to-10 list plus an appended WEEKLY_DIGEST.md entry. Self-schedules a weekly cron; also fires manually. Use whenever the user asks to "run roadmap pulse", "rebalance the backlog", "score open tasks", "what should I focus on this week", "is my roadmap up to date", or any phrasing implying both honesty-checking the planning docs AND deciding what's next. Triggers even on passive cues like "feels like the backlog needs a refresh" or "I'm not sure what to focus on" — even when the user doesn't say "skill" or "pulse" explicitly. Output is a sourced prioritized list with RICE scores + reusability/strategic-fit/dependency-unblock overlays, plus a dated WEEKLY_DIGEST.md section that builds a decision history.