How I build software with AI agents.

The reusable workflows I use to turn a problem into working, reviewed software. Each skill has a job, a clear input, and a useful output. I keep the judgment; the agent helps carry out the work.

From a problem to a pull request

Choose where you are starting. Select a skill to see what it does and what comes next.

Explore the idea with brainstorm, then define behavior, design, and a plan. Implement directly or create issues for later. take is reserved for issues created outside this workflow.

Understand & decide

Need a backlog? Take this detour after planning.

These issues already contain the specifications. Continue to code-tdd without take.

Otherwise, go straight from plan to code-tdd.
Implement & verify

Review found something? Evaluate it, apply agreed corrections, then recheck the affected behavior.

Back to the flow
brainstorm

Explore the idea

Start with a rough idea, question, or concern. Explore the problem, examine alternatives and tradeoffs, and clarify a useful direction before committing to requirements.

Read the skill
What you get
Shared understanding, a direction or decision, and the uncertainties that still matter.
What happens next
When the idea is ready to become development work, use dev to coordinate requirements, design-spec, and plan.

dev coordinates the development stages.

dev coordinates the stages the work needs and passes explicit inputs and outputs between workers. I can ask for the whole flow, stop at requirements, or continue from an existing plan. It reuses sufficient earlier work. The map shows the responsibilities inside that flow, not a checklist to repeat on every change.

$dev Add CSV export to the filtered customer list. Keep existing permissions and include review.

Use only the flow the work needs

An idea that still needs decisions

Explore the idea with brainstorm, then let dev coordinate the necessary stages.

$brainstorm How could we make customer data easier to export?
$dev Continue from this direction through requirements, implementation, and review.

A plan that becomes a backlog

Create issues only when you want tracked outcomes. Select one later and carry its existing specifications forward.

$create-issues docs/plans/customer-export.md
$next-issue
$dev Implement the selected issue using its existing specifications.

An issue created elsewhere

Use take to understand an issue that did not come from create-issues. Fill gaps before implementing.

$take Explain this external issue and what should change.
$dev Complete the missing specifications, then implement and review.

A change that is already specified

Skip completed preparation. After review, optionally record an interactive change with record-pr and include the media in create-pr.

$dev Implement this plan with code-tdd, then run review-run.
$record-pr Record a walkthrough of the changed UI.
$create-pr Prepare the PR with the walkthrough and validation.

Examples use $ in Codex. In Claude Code, use / before the skill name.

Small change? Use mini-plan.

For a bounded improvement, mini-plan keeps requirements, design decisions, execution steps, and verification in one short document. The amount of ceremony should follow the uncertainty and risk.

Make the skills yours

The public repository contains the workflows. Your private profile tells them where your projects and context live. Your agent client supplies access to connected services.

Install and answer the questions

Use Python 3.10 or newer. Choose codex, claude, or both. Setup asks for your name, language, timezone, private context directory, and selected optional modules. Enable work for repository defaults and integrations for optional service configuration.

git clone https://github.com/lucastamoios/skills.git
cd skills
python3 install.py --client codex

Configure your private profile

Answers are saved in ~/.config/agent-skills/config.json. Set SKILLS_CONFIG to use a different file. Run skill-config in a new agent session to update these settings later.

work.github_repository · work.github_login
Your repository and GitHub identity.
work.issue_prefix · work.task_destinations
Your issue prefix and routing to Linear or private task notes.
integrations
Optional service contracts, procedures, and credential references.
execution.model_policy
A path to your private worker model assignments. Coordinators need these assignments or explicit choices.

Connect Linear, GitHub, and other tools

Connect or configure the service in the agent client where you run the skills. For Linear, the client needs authenticated tools that can read issues and, when requested, update or create them. The skill can then resolve your user, team, and real workflow statuses from the connected service.

The questionnaire does not install Linear, authenticate accounts, or grant permissions. There is no required Linear API key field in the profile. Keep token values in a vault or injected environment; store only references in configuration.

$skill-config Help me configure my repository, issue prefix, task destination, Linear integration, and development/review model policy. Ask only for the missing settings.
Read the installation guide

Explore the full collection

Development is one part of the collection. The repository also includes skills for explanation, conceptual analysis, knowledge, business, and content.

GitHub