aaronkendell/dev

Test API endpoints in a workspace using OpenAPI specs, auth tokens from Infisical, and self-healing debugging for common failures

Provision a new app across Cloudflare, Infisical, Neon, Fly, Sentry, Grafana, Upstash and the rest — decide which vendors it needs, create what's scriptable, and print the manual checklist

Stand up a NEW app repo from project-template — config, catalog, design pack, generated DESIGN.md, CI wiring — and hand off to app-provision for vendors. Use when the user says "new app", "scaffold an app", "create the <name> repo", "bootstrap <name>", or asks what a new app needs. Covers the repo side only; app-provision covers Cloudflare/Infisical/Neon/Fly/etc.

App Store Connect as code for any bokendell mobile app. Use when touching store metadata, screenshots, TestFlight, review submission or a mobile release; it names the folder, scripts, secrets and authentication every app shares, and defers the command detail to the installed asc binary's own help.

Run every CI check on your own machine before pushing, so GitHub Actions minutes are spent confirming rather than discovering. Use when a PR is about to be opened, when CI just failed on something a local run would have caught, or when deciding which workflows should be allowed to run on a draft PR.

Retrospective on a finished piece of work — where the loop deviated from its skills, what was done by hand that no skill covers, which playbook command lied — emitted as improvement signals per improve. Use after a merge, or when a session ends with something worth counting. Recommends only; applies nothing.

How to run a bokendell app repo inside a cloud VM — a Cursor Cloud Agent or a Claude Code cloud session (vs swarm workspace dev on a laptop). Use when starting servers/apps (API, admin, web, mobile/Metro, Inngest, workers, MCP servers), deciding which services a task needs, wiring Infisical/swarm in the sandbox, or exposing something to a phone/browser over a tunnel. On a laptop keep using swarm workspace dev.

Run the local gate — the lefthook / CI-equivalent checks via dev:ci-local and the repo's verification skill — then stage deliberately and commit. Only runs when the user asks to commit; never auto-invoked by another stage. Refuses to commit on a red gate.

Surface what already governs a piece of work before designing it — the repo's docs map, decision records, product journeys, and the memories that apply. Use at plan time (there is no file yet for auto-loading to anchor on) and at review time to load the rubric. Reads only.

Auto-loads the correct pattern docs based on what's being worked on. Triggers when working on golf, portfolio, or hive apps — during planning, development, testing, or code review. Ensures DDD, API, frontend, mobile, and testing patterns are always in context. Use proactively whenever implementing, reviewing, or planning features.

Voice and copy for every word a user reads. Use when writing or reviewing UI strings, empty and error states, buttons, toasts, push notifications, emails, onboarding, marketing headlines, or App Store text. Also use when the user says copy sounds "generic", "AI-written", "off-brand", "slop", or asks about tone, humor, em dashes, microcopy, or naming a button. Companion to the design skill: design owns pixels, this owns words.

Unified design orchestrator for every app (golf/Bagman, portfolio, hive, keepings). Use when building UI components, designing screens, creating mocks, reviewing design compliance, updating design tokens, exploring color palettes or typography, implementing animations/motion, or doing any visual/frontend work. Also use when the user mentions "design system", "make it look better", "fix the UI", "design review", "polish this", "update the theme", or references any app's visual style. Also use when the user says design studio, golf design studio, studio sketch, flow sketch, primitive library, or app-specific design studio. Triggers proactively on ANY frontend/mobile file changes to ensure design compliance. This skill replaces golf-ui and is the primary studio router for every app, including golf.

Inbox + address workflow for design annotations. Lists every open annotation across an app's flows (or filters to one flow), then walks through each — reading the comment + the source file + the flow's decisions.md, making the fix, and posting a reply that marks the annotation addressed. Triggers when the user says: "check comments", "address feedback", "any new design comments", "let's go through the feedback", or invokes /design-comments / /design-address.

Multi-agent taste review for a design flow. Composes /design-verify (structural / HARD-RULES evidence) with parallel passes through /taste-skill, /impeccable:impeccable, /ui-ux-pro-max, and /huashu-design. Each agent writes its findings to swarm-api as annotations tagged with its own author (agent:taste, agent:impeccable, agent:ui-ux-pro-max, agent:huashu). The pin layer's agent badge surfaces who said what. Triggers when the user says: "review the flow", "do a design review", "what would taste-skill say about this", "have impeccable polish-check it", or invokes /design-review.

Visual verifier for design studios. Captures rendered screenshots + DOM + computed styles per artboard via the playwright-cli skill, then audits each one against the flow's decisions.md, packages/ui/HARD-RULES.md, and the design tokens. Findings POST to swarm-api as annotations tagged origin: "agent:verify". Triggers when the user says: "verify the studio", "check this flow's design", "run the visual verifier", "make sure the artboards match my decisions", or invokes /design verify. Use INSIDE Claude Code only — it authenticates as the shared Claude agent identity (/apps/swarm/agents in Infisical) so findings are attributed to the agent persona, not the user.

Compatibility entry point. /dev, /dev build, /dev <LINEAR-ID> route to drive; /dev research → dev-research; /dev plan → dev-plan; /dev ship → commit-change, open-pr. Use at the start of any substantive dev task.

Retired alias. /dev build <LINEAR-ID> now routes to drive, which detects the stage and runs draft-approach → implement → review-change → map-paths/exercise-paths → commit-change → open-pr. Kept so old muscle memory still lands somewhere.

Use when the user asks for the dev plan workflow to break a design document into Linear projects and issues. Triggers on /dev plan <app> [initiative-path] or when the user says "break this down", "create issues", "plan this project", "sync to Linear". Reads the design doc, decomposes into projects and issues, writes local markdown files, and syncs to Linear.

Use when the user asks for the dev research workflow to create PRDs, database designs, architecture docs, API specs, and UX mocks for a new feature or app. Triggers on /dev research <app> <idea> or when the user says "research this", "design this feature", "create a PRD", "brainstorm this idea". Combines codebase scanning, brainstorming dialogue, and design document generation into one phase.

Retired alias. "Ship it" is now three explicit, separately gated steps: commit-change (local gate then commit), open-pr (draft; --ready undrafts and merges on explicit asks → Linear Done → docs distill).

Hand a task to another repo's cloud agent by firing that repo's saved routine with the task text. Use when a session needs a change in another repo (Bagman needs core), when a shipped PR needs landing, or when an orchestrator kicks off work. Never for releases.

The bokendell docs contract: where docs live, how to find the right one cheaply, when to update them, and how to distill planning docs on ship. Use when the user says "update the docs", "where does this doc go", "archive this project", "distill", "docs map", mentions MAP.md, or whenever creating/restructuring documentation in any bokendell repo. Also use when the docs-freshness hook reports stale owning docs. Spec lives at docs/planning/docs-context-layer/design.md in the golf repo.

Grill the requirement, load constraints, write the technical notes with receipts, and lock a plan into the branch's plan file. Use for any medium/high-band work before implementing. Refuses to lock while an open question remains.

Drive a piece of work from wherever it is to shipped. Takes a Linear issue, a rough idea, a doc path, or nothing. Detects the stage the work is at, sets the ceremony band, and does the next thing. The main entry point — /drive (and /dev) route here. One-off fixes need no ticket; the branch slug is the key.

Run the map — tests, API calls, jobs, DB reads, Maestro flows — and report hit/miss per target with the misses left in the denominator. Hands device-only targets to the user as a checklist. The QA stage; wraps dev:verification and the repo's verification skill.

Post agent feedback to the repo's living feedback ticket at the end of a session, including what you could not do but wanted to, so tooling and skills improve from use. Use when a session hit a wall, a skill was wrong, a tool answered badly, or something was missing.

Stop GitHub Actions from running on work nobody has looked at yet — draft-PR guards, Renovate draft PRs, concurrency groups and path filters. Use when adding or editing a workflow, when Actions spend is too high, when an agent's exploratory PRs are firing full pipelines, or when deciding whether to self-host runners.

Compatibility alias for the name people still type. The golf design pack now lives in the golf repo at .claude/skills/design/SKILL.md and loads automatically there; the cross-app method lives in dev:design. This file only redirects.

The shared improvement contract — what a signal looks like and where it goes — so a gap seen by close-out, weekly-skill-review, the monthly audit or a human is counted once and fixed once. Backed by skill-watch's store. Read by every consumer; owned by none.

Enumerate every reachable path a change touches — unit, integration, API contract, authz both directions, realtime, jobs, DB, UI journey, device-only — and what already covers each. Documents only; runs nothing. Use before exercise-paths; the split is what keeps gaps visible.

Teach the dev loop how a repo works by writing (or repairing) its playbooks — the repo-local skills that drive, map-paths, exercise-paths, commit-change and use-worktree read. Use when entering a repo with no .claude/skills/verification, or when a primitive says a playbook command is missing or wrong.

Start a piece of work — branch (in a worktree by default), Linear moved to In Progress, plan file seeded for medium+ bands. Use when beginning a ticket or a named one-off. No ticket required; a slug is enough.

Open a DRAFT pull request from the current branch so GitHub Actions do not run yet. Body from the plan file, review notes and exercise-paths output — never from memory. Undrafts (--ready) and merges only on an explicit ask, each a separate yes.

Find OpenAPI specs for every API in the monorepo — both on-disk JSON files and live URLs served by each API

One tick of the agent orchestrator. Read the board (Linear + open PRs), decide what to dispatch, fire it, post one digest, write the journal. Use when running the dispatch loop by hand from a local session, and later as the prompt of the scheduled orchestrator routine.

Capture a non-obvious decision or trap while the reasoning is still available — into docs/decisions (durable), the plan file (branch-scoped), or auto-memory (cross-session). Use when a "we chose X over Y because Z" moment happens mid-work. Never edits a durable doc's status; that is the docs contract's job.

Write exercise-paths results where they persist — the plan file's Gate, the PR body's "How it was verified", a Linear comment, and for area passes docs/qa/<area>.md. Reporting only; runs nothing.

Review files against the project's pattern docs (DDD service/repo, Hono+oRPC API, Awilix DI, testing layers, frontend container/component split, mobile, design app, design studio annotations, CLI, barrels, container-images, AI/Mastra, remote tunnels, etc.) and fix violations so the code follows the patterns exactly. Use this skill whenever the user says "review", "check against patterns", "make this follow the patterns", "review this PR", "fix pattern violations", "audit this against our conventions", "is this DDD-compliant", "does this match the api.md pattern", "review my changes", or invokes a /review style command — even if they don't name a specific pattern doc. Trigger when the conversation mentions specific files to audit OR when the user describes a code-quality / convention-compliance pass with no specific framework named (the skill maps file globs → pattern docs automatically). Use this *instead* of the patterns-reviewer agent for fast, on-demand, interactive review (the agent is for batch/background review during the build phase).

Review a change as the only reviewer it will get — patterns via dev:review, correctness via the built-in /code-review at an explicitly chosen level, security via /security-review when warranted. Proposes the exact invocation first. Use before QA on your own branch, and in --agent-branch mode on anything a cloud agent produced.

Use when implementing a feature and noticing changes that may be outside the planned scope. Triggers during execution when edits touch different domains, add unplanned features, or modify infrastructure not in the issue spec. Also use when reviewing a diff that seems larger than expected. Helps prevent scope creep during build phase.

How config and secrets reach running code — the zod config boundary, the three injection paths (local, CI, Fly runtime), why a value in Infisical is not yet a value on a machine, and the NEXT_PUBLIC build-time exception. Use when adding a config value, rotating a key, or debugging "it's set but the app can't see it".

Create, list, and prune git worktrees under <repo>/.worktrees/<name>, and bind a swarm workspace (own Neon branch + tunnels) to one when the work needs a live stack. Owns the protocol; the repo's dev:workspace and verification skills own the commands.

The verification loop for any app change — pick the cheapest check that would catch this change breaking, run it, and LOOK at the result. Triggers when the user says "verify this", "prove it works", "run the verification loop", after any non-trivial change, and ALWAYS before declaring a feature done.

Weekly pass over the improvement store — group by recurrence, propose concrete diffs for anything at threshold, never apply. Delegates to skill-watch:skill-watch-review for the telemetry read and skill-watch:skill-watch-promote for the hardest-layer triage.

Manage dev workspaces — create, stop, start, status, URLs. Use when the user mentions "workspace", "start servers", "stop servers", "which workspace", "what port", or "show URLs".

Work productively inside a workspace — know ports, URLs, env vars, CLI commands, health checks

Bootstrap a workspace from scratch — full lifecycle from zero to running dev environment, portable to cloud workers