Open-source tools by stark AI
Agent Skills: reusable working methods for Codex
Agent Skills make proven workflows installable, from clarifying fuzzy requirements to checking architecture and mapping code structurally. The public catalog remains the current source for versions and installation.
A skill describes the method, not only the goal
An Agent Skill is a versioned set of instructions for recurring work. It can define which sources to read first, which decisions need human confirmation, which files matter, and how to validate the result.
That turns a general prompt into an inspectable working method. Teams can reuse, review, and deliberately improve the same workflow instead of starting from scratch each time.
When a skill creates real value
01
Recurring work
A clear method for frequent tasks reduces missed steps and inconsistent results.
02
Important boundaries
Skill instructions can embed architecture, security, and approval rules in the workflow.
03
Verifiable outcomes
A strong skill identifies concrete artifacts, validation, and an unambiguous completion point.
04
Deliberate selection
The catalog separates promoted skills from experimental candidates and exposes provenance.
Choose a skill that fits the task
Not every skill belongs in every workflow. Start with the problem, review the skill boundaries, and install only the working method that covers a concrete missing step.
- 01
Identify the problem class
Is the gap clarification, architecture, impact analysis, or another clearly named task?
- 02
Review a promoted skill
Read its description, expected inputs, outputs, boundaries, and current version in the public catalog.
- 03
Install it selectively
Bring only the required skill from the public repository into the local agent environment.
- 04
Invoke it deliberately
Use the skill for the relevant phase and review its output before continuing.
- 05
Finish with a verifier
Repository-owned checks and human judgment remain binding even with a structured skill.
01
Codex Spec Interviewer
This skill turns a fuzzy coding request into a verified implementation specification. It is useful when the objective, boundaries, acceptance criteria, or sources are not dependable yet.
- Useful before features and ambiguous changes.
- Makes assumptions and open decisions visible.
- Creates a reliable hand-off before execution.
02
Architecture Compass
Architecture Compass aligns implementation and review with ADRs, stack rules, runtime boundaries, and accepted examples. It helps place new files and responsibilities in the owning layer.
- Useful before architecture-relevant features and refactors.
- Separates target rules from current repository drift.
- Requires explicit placement and focused validation.
03
CodeGraph with ast-grep
This skill uses structural search and code graphs to understand callers, shapes, and likely impact. It is stronger than plain text search when syntax and relationships matter.
- Useful for test fixes, CI problems, and focused reviews.
- Finds related structures and dependent callers.
- Supports a smaller, better justified change.
04
A useful integration with LoopLatch
LoopLatch recommends suitable skills for each template and explains their role. The recommendation is a hand-off: customers deliberately install and use a skill before the loop. The prompt and ZIP remain unchanged so the generated harness stays understandable and inspectable.
Catalog, sources, and application
Common questions about Agent Skills
How is a skill different from a prompt?
A prompt often describes one task. A skill is an installable, versioned working method with sources, boundaries, steps, and validation for a recurring problem class.
Do I need to install every skill?
No. Install only skills that address a concrete gap in the current workflow. The catalog is a selection surface, not a required bundle.
Is every promoted skill automatically suitable for every repository?
No. Promoted means the skill has passed the public quality and release path. Repository rules, required tools, and the specific scope still need to fit.
Does LoopLatch alter my prompt when it recommends a skill?
No. The recommendation is separate guidance. Installation IDs and commands are not automatically written into the generated harness or ZIP.
Start with a suitable skill, not a larger prompt
Choose the smallest working method that covers your missing step, then combine it with the validations owned by your repository.