.agents

Agent skills

Plan, review and ship with skills for T3 Code, Codex and Claude Code.

Install from your project folder with Node.js 22.20+.

Plan and design

  • think

    Turn an idea into a brief. Answer key questions and get a second opinion before you build.

    npx skills add seankoji-com/.agents --skill think

    Choose your agents when prompted.

    Try asking Use think on adding replays to the game before we build anything.

    Supporting skills include review, choose-design, and raise-issue.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • challenge

    Get a second opinion on a proposal you already have. Find weak assumptions, a simpler option and evidence that would help you decide.

    npx skills add seankoji-com/.agents --skill challenge

    Choose your agents when prompted.

    Try asking Challenge this plan to split checkout into its own service.

    Supporting skills include review.

    Close install instructions
  • choose-design

    Pick a design before building. Compare independent mockups for a new screen, redesign or unresolved UI choice.

    npx skills add seankoji-com/.agents --skill choose-design

    Choose your agents when prompted.

    Try asking Show me design options for the new settings page.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • focus-group

    Explore how fictional participants might respond to your product. Get a clearly labelled synthetic report and ideas to test with real users.

    npx skills add seankoji-com/.agents --skill focus-group

    Choose your agents when prompted.

    Try asking /focus-group on our pricing page with five first-time visitors.

    Invoke this skill explicitly, for example with /focus-group.

    Supporting skills include review.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions

Build and deliver

  • ship

    Hand off a change through implementation, tests, independent review and authorised delivery. Get the result, evidence and a short retrospective.

    npx skills add seankoji-com/.agents --skill ship

    Choose your agents when prompted.

    Try asking Ship the agreed replay feature. Verify it in the browser first.

    Supporting skills include think, choose-design, plan-report, implement, review-loop, review, raise-pr, babysitter, raise-issue, and github-writing.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • plan-report

    Get a timeline of the plan, work underway and decisions waiting on you. Compare actual build and review work with the original estimates.

    npx skills add seankoji-com/.agents --skill plan-report

    Choose your agents when prompted.

    Try asking Use plan-report to show this coordinated run and the decisions waiting on me.

    Close install instructions
  • implement

    Requires T3 Code

    Give a coding job to a lead in a separate T3 worktree. Get reviewed commits back in the main thread for integration.

    npx skills add seankoji-com/.agents --skill implement

    Choose your agents when prompted.

    Try asking Implement the export feature from this brief.

    Supporting skills include plan-report, review-loop, review, and raise-issue.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • raise-pr

    Open or update a pull request with its required checks, review details and screenshots where relevant.

    npx skills add seankoji-com/.agents --skill raise-pr

    Choose your agents when prompted.

    Try asking Open a PR for this branch.

    Supporting skills include review and github-writing.

    Close install instructions
  • babysitter

    Have the agent repair, watch and merge an open pull request through its checks and reviews. Say “don’t merge” or “read-only” to limit the work.

    npx skills add seankoji-com/.agents --skill babysitter

    Choose your agents when prompted.

    Try asking Babysit PR 42 through to merge.

    Supporting skills include review, raise-pr, and github-writing.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • backlog-burn

    Requires T3 Code

    Clear a repository’s issues and pull requests. Close obsolete work with evidence and get ready fixes merged. Name an organisation to run it per repository.

    npx skills add seankoji-com/.agents --skill backlog-burn

    Choose your agents when prompted.

    Try asking Clear the backlog and the open PRs. Ask me what you need to unblock, then implement and merge the ready work.

    Supporting skills include ship, plan-report, babysitter, and implement.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • raise-issue

    Requires T3 Code

    Draft or file a GitHub issue in a background T3 task. Check for duplicates, attach evidence and verify the link after posting.

    npx skills add seankoji-com/.agents --skill raise-issue

    Choose your agents when prompted.

    Try asking Raise an issue for the flaky login test with the failing logs.

    Supporting skills include github-writing.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions

Review and improve

  • review

    Get an independent review of a plan, change or release. The reviewer reports ranked findings with evidence and leaves the work untouched.

    npx skills add seankoji-com/.agents --skill review

    Choose your agents when prompted.

    Try asking Use review to check this branch against the brief. Do not edit files.

    Supporting skills include implement, raise-issue, and github-writing.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • review-loop

    Build or fix a change in rounds. Independent reviews guide repairs until the score and checks pass or the work stops improving.

    npx skills add seankoji-com/.agents --skill review-loop

    Choose your agents when prompted.

    Try asking Use review-loop on the search fix until review scores it 9 or higher.

    Supporting skills include review, implement, and raise-pr.

    Close install instructions
  • hillclimb

    Improve a product, workflow or draft using scored feedback from judges you approve. See each round’s evidence. UI judges use the running interface.

    npx skills add seankoji-com/.agents --skill hillclimb

    Choose your agents when prompted.

    Try asking Run hillclimb on the onboarding flow until it scores 8 out of 10.

    Supporting skills include review, raise-pr, and raise-issue.

    On T3 Code, also install agent-tuneup for the permission check.

    Close install instructions
  • test-writer

    Write or review regression tests for failures that matter. Weigh coverage against flakiness, maintenance and CI time.

    npx skills add seankoji-com/.agents --skill test-writer

    Choose your agents when prompted.

    Try asking Use test-writer to add regression tests for the date parsing bug.

    Close install instructions

Write on GitHub

  • github-writing

    Add a byline to AI-written GitHub prose. Credit the writing model and the verified operator in issues, PRs, replies and release notes.

    npx skills add seankoji-com/.agents --skill github-writing

    Choose your agents when prompted.

    Try asking Use github-writing to draft the release notes.

    Close install instructions

Fix agent setup

  • agent-tuneup

    Diagnose missing skills, permissions and connections. Trace what the agent does back to the setting that controls it.

    npx skills add seankoji-com/.agents --skill agent-tuneup

    Choose your agents when prompted.

    Try asking Use agent-tuneup: why doesn't my agent see the review skill?

    Close install instructions

Install and use a skill

  1. Run the install command

    Open Install next to a skill. Run the command in your project terminal and choose agents when prompted.

  2. Ask for the work

    For example, ask Use think on this plan before we build it or run /think if your agent supports slash commands.

  3. Check the result

    Your agent follows the skill's instructions and reports the outcome, with any decisions or checks still open.

Made for T3 Code

Use T3 Code to run agents from several providers in one app. Skills use T3 for separate worktrees, background tasks and pull request watches.

A second opinion from another model

review, think and challenge get a separate opinion. Written-code reviews use a different model family from the code authors. Read the rationale.

Big jobs in their own workspace

implement does the coding in a separate T3 worktree, so your main conversation stays free for decisions.

Work that carries on in the background

raise-issue files issues in a background task while your change keeps moving.

Pull request watches

babysitter uses T3's pull request watch to resume work when checks or reviews change.

Progress and designs you can see

ship, implement and review-loop use T3's Tasks row when the host exposes it. plan-report draws a timeline, and choose-design shows mockups side by side.

Using another agent

Codex, Claude Code and other agents can use skills whose required tools they provide. Look for the T3 Code requirement in the catalog.

Why another model family checks the work

Models can favour their own writing and repeat mistakes made by related models. Using another family for code review gives the change an independent check.

  • GPT-4 and Llama 2 tended to prefer text they could recognise as their own. Another study linked GPT-4's preference to how familiar the text was. Panickssery, Bowman and Feng, 2024 and Wataoka, Takahashi and Ri, 2024.

  • When two models answered incorrectly, they chose the same wrong answer 60% of the time on one benchmark. A shared provider or architecture made that more likely. Kim et al., ICML 2025.

  • Models sometimes made their reasoning worse when asked to correct it without outside feedback. Huang et al., ICLR 2024.

  • A panel drawn from several model families judged more accurately and with less bias than a single GPT-4 judge across six datasets in one study. Verga et al., 2024.

The studies tested AI judges and reasoning benchmarks. Read how the library applies those findings to code review in the repository README.

What you need

T3 Code, or another coding agent

On T3, install agent-tuneup when a skill calls for its permission check. For how skills choose delegate routes and check subscription capacity, see the workflow guide.

Python for skill helpers

Some skills run helpers that need Python 3.10 or later. See the Skills CLI setup guide for installation options.

Tools for the job

Installation adds the instructions. Review, browser checks and delivery use the models, tools and permissions available to your agent.