Skills

Save the goal, progress, evidence, and next action between rounds. Independent audits require fresh agents.

Each round gets a fresh builder, an inspector and a panel of judges, with written verdicts and a run journal of every agent's input and output. Claude Code only; Codex sessions use Long Horizon.

Parallel candidates tackle the same brief. A fresh judge chooses a base, then useful ideas from the other candidates are grafted in.

Independent context and inspected evidence support actionable findings. Requires exposed agents or an authenticated Codex CLI.

Retrieve relevant project facts and pitfalls with evidence pointers. Save durable learning only within the authorized scope.

Use rendered evidence, independent criticism, and verified repairs for a demanding deliverable.

Each round starts from a baseline, makes one focused change, and checks the result with an independent review. Covers frame rate, loading, latency, throughput and resource use.

Capture baseline and treatment under comparable conditions. Return VERIFIED, NOT VERIFIED, or INCONCLUSIVE.

Explore the product, interface, or architecture choices that materially affect the result.

Break a multi-step task into concrete work, dependencies, and verification.

Fresh agents decompose the problem, audit assumptions, recombine what survives, and test it against reality.

Create a self-contained prompt for another agent or a fresh session. The prompt is the deliverable.

Write a self-contained prompt with exact scope and falsifiable checks. This skill drafts the audit rather than running it.

On explicit invocation, use independent scrutiny to test whether the recent change should proceed.

Explicit invocation asks for an independent challenge to the previous recommendation.

Run the Codex CLI review workflow with gpt-6-astra at medium reasoning, then verify each finding. Requires the CLI and model access.

Run codex exec review with gpt-5.6-sol at high reasoning over a PR, branch, commit, or uncommitted diff. Verify each finding.

Use the Claude CLI for a separate review and verify the reported findings. Requires an authenticated Claude CLI.

Build an isolated visual or interaction experiment with live controls, then port the chosen result.

Choose a direction from the real material, test it in the intended medium, and refine the details people encounter.

Draft or revise site copy, docs, release notes, and other public text while preserving facts and voice.

Author or revise skill instructions and verify their behavior before use. Follow the repository install contract when adding them.

Research relevant UI skills and synthesize a focused skill around the project’s actual design needs.

Inspect PR checks, diagnose failures, and repeat verification. Repository access and authorized pushes are required for remote changes.

Review the session for corrections, wasted work, and missing guidance. Make narrow improvements supported by what happened.

Inspect kernels, skills, and connectors for avoidable token cost. Propagate generic improvements to the starter when requested.

Use project transcripts and direct questions to draft a personal mode skill.

Use deliberate diagnosis, verification, and handoff for layered tasks, uncertain causes, and repeated failures.

Restate the last response without jargon while preserving its meaning.

Compress conversational prose while retaining technical accuracy. Code, commands, and deliverables keep their normal form.

Read the project’s facts, fill its setup gaps, and configure only what the project needs.

Mirror the external repository privately, check for private data, and overlay the starter setup. Requires the relevant account access.

Register a repository skill so future Claude Code and Codex sessions can find it.

Compare template drift, pull starter updates, or send generic project improvements back when requested.

Use a shared append-only Desktop HTML hub to track session work and avoid collisions. This workflow targets Claude Code sessions.

Explicitly authorize session-wide commit, push, pull request, and merge behavior. Requires repository access; this mode has remote side effects.