Agent skills maintained by @marcelsud. Each directory contains a SKILL.md entrypoint and any supporting prose references. The workflows are independent of agent harnesses: they describe outcomes, decisions, and verification without requiring a particular tool API, command syntax, installation path, or runtime configuration.
Grouped by their role in the software development lifecycle.
| Skill | Purpose |
|---|---|
prep-dev |
Examine the premise and clarify implementation through a short interview. |
fusion |
Run two user-selected models on one prompt and reconcile the answers. |
| Skill | Purpose |
|---|---|
unit-tests |
Evaluate test relevance and protect business use cases and domain invariants in any language. |
| Skill | Purpose |
|---|---|
spec-gap |
Compare code with product requirements; score coverage and identify gaps and extras. |
sheriff |
Review code changes and triage issues by evidence and risk. |
simplify-code |
Simplify working code for clarity while preserving its exact behavior. |
prep-pr |
Assess implemented changes and prepare an evidence-backed review handoff. |
| Skill | Purpose |
|---|---|
warden |
Validate intent, review, tests, documentation, and lint; then deliver within the authorized scope. |
| Skill | Purpose |
|---|---|
docs-writer |
Write and improve open source documentation with runnable examples, progressive explanations, and precise references. |
gatekeeper |
Acceptance criteria, evidence, and the Depth Tree for substantial tasks. Fork of Leonxlnx/unlazy. |
memory-sql |
Persist and selectively recall project facts, decisions, tasks, and evidence through SQLite. |
Load the chosen skill through your environment's supported skill mechanism, or provide its entrypoint directly. Keep linked reference documents available and read them when the workflow calls for them. The directory name identifies the skill.
Standard frontmatter supplies the skill's name and description, with license and attribution where applicable. The remaining content describes the workflow in prose. There are no bundled scripts, database schemas, hooks, or harness-specific metadata files.
Use the execution capabilities available in the current environment. Some tasks still need capabilities inherent to their purpose: memory-sql uses SQLite, fusion needs analyses from the two selected models, and a required independent review needs a separate reviewer. Report unavailable capabilities without claiming the corresponding work or verification passed.
Before testing an edit, confirm which copy of the skill the environment actually loads. A separately installed copy can hide changes made in this checkout. Use the environment's supported way to load or refresh local instructions and avoid divergent copies of the same skill.
Preserve each workflow's intent, evidence requirements, and authorization boundaries when editing it. Describe capability requirements in ordinary language and leave their implementation to the environment. Keep references current and retain applicable license attribution.