Development

Describe what you want to build. The included Development instructions guide your AI coding tool through planning, building, review and verification. You supply the tool, account and model access; CoreMVP does not provide a managed team.

Use $development to add an editable description to Projects. Owners and
admins can change it; members can read it. Keep the work local, preserve the
dashboard's accepted style, and verify the saved value and permission checks.

Open the purchased repository in Codex or Claude Code and give it your outcome. The example uses Codex $development; use /development in Claude Code. Lead plans in your main conversation, Review challenges the plan, Code implements and runs the checks, and Review inspects the source and evidence. Lead resolves findings and delivers the result. You do not need to select skills or manage agents.

Describe the result

The initial workflows cover a feature, a reproducible bug and an interface improvement. These are example requests, not recorded execution transcripts:

Include a known constraint, affected user or failure when it matters. Lead checks source and accepted decisions before asking material questions. Every question comes with its executive recommendation, why and the practical tradeoff, followed by one clear decision for you. It records scope, acceptance/rejection cases, dependencies and verification before implementation.

AGENTS.md routes complete requests to .agents/skills/development/SKILL.md. Four core skills divide the work:

The phase procedures are .agents/workflows/proposal.md, challenge.md, implementation.md and ai-code-review.md. $development coordinates them for complete delivery. You can request one phase only; a proposal, challenge or review does not automatically start implementation. Lead keeps decisions and phase outputs in one ignored task record.

The eight separate domain skills are frontend, backend, data, auth, billing, product, local-development and writing. Lead loads the guidance needed for your change. Marketing uses the same Lead/Code/Review actors for content, with Writing, Taste and SEO expertise. Ordinary marketing does not run the four engineering phases; technical changes still use Development.

Judge an unsettled visual direction

For a new or unsettled interface direction, expect actual screenshots of inspected references, the current application where relevant, and a rendered proposed desktop/mobile mockup. The lead explains the information, hierarchy and interaction it recommends and labels illustrative data or unwired controls.

You judge that direction before application implementation. Once you accept it, the lead continues bounded work and small fixes under your existing instruction to implement, without another request to continue. A rejected direction returns to the information and composition; an existing approved interface needs only the discussion relevant to the change.

.agents/skills/frontend/SKILL.md owns this process. Its references/ui-references.md directory groups shadcn and other sources by useful design problems. A reference must fit your product; the directory is a starting point for inspection, not a prescribed dashboard layout. Taste contributes consequential visual judgment and pertinent project feedback while Frontend retains ownership of the interface process.

Check the finished change

The lead finishes with the working behavior, review findings and their resolution, the checks that ran, and any unverified boundary. It keeps a short local progress record so an interruption can resume from the accepted decisions and next action. Local work, commits, publishing and deployment are reported separately.

Verification follows the changed behavior. For the included persisted Projects application, the local command is:

./coremvp e2e projects

Start the app through Getting Started before running it. The suite uses disposable local accounts, real Supabase Auth, app routes and Postgres. Organization memberships include test fixtures; this suite does not complete a Stripe checkout. Your new feature needs checks for its own acceptance and rejection cases, rather than assuming this existing suite covers it.

Included guidance

Ten reusable skills and four engineering procedures are included in this checkout alongside five application skills: backend, data, auth, billing and local development. AGENTS.md supplies this foundation's architecture, permissions, component choices and SEO file map; generic skills do not replace it.

When updating, preserve your instructions, customized files and learned Taste judgment.

Client setup and limits

Codex discovers canonical .agents/skills and uses the existing .codex/agents/code-agent.toml and review-agent.toml definitions. Claude Code reads CLAUDE.md -> AGENTS.md, a relative symlink, the individual .claude/skills symlinks, and .claude/agents/code-agent.md / review-agent.md. Both keep Lead in the main conversation and inherit the parent's model. Claude definitions preload their canonical role skill; Code uses normal client tools/permissions and Review allows only Read, Grep and Glob.

Use Codex $development or Claude Code /development. Client version, trust and settings affect discovery and matching. A same-named personal or inherited skill can be exposed instead of this repository's copy; if a listed description differs from .agents/skills/<name>/SKILL.md, have your tool read that file by path. If Git/your platform materializes symlinks as text or native types are unavailable, load .agents/skills/development/SKILL.md explicitly and assign available agents the canonical Code/Review roles. Without delegation, sequential role switches are self-review; independent review did not run.

Other tools can consume the Agent Skills format, but it does not establish their native discovery, permissions or orchestration. See Codex Skills, Codex Subagents, Claude Code Skills, Claude Code Subagents and Claude memory imports for client contracts.

Review never writes files, runs checks, operates a runtime/browser/provider or delegates. Missing proof returns to Code. In Codex, Review inherits the session's permissions and stays read-only by following its role instructions. Claude's allowlist excludes Bash, Write, Edit, Skill, Agent and MCP tools. Configuration recognition is separate from model selection and end-to-end delivery. No hosted agent service is included. Publishing, deployment, paid provider calls and production changes require explicit authorization.

What's next