tiny-skills

Self-contained, production-ready agent skills repository.

Each directory under skills/ is an independent agent. Skills are grouped by class for discovery, while remaining fully independent. Agents do not depend on, import, or coordinate with other agents in this repository. Every agent follows the Agent Skill standard and can be installed directly with:

npx skills add smota/tiny-skills --skill <skill-name>

Skill catalog

Skills listed here. Each skill's own SKILL.md is authoritative for detailed usage.

Class	Skill	Purpose	Prerequisites	Install
software-engineering	empirical-refinement-loop	Run bounded, reviewed refinement cycles with comparable evidence and repository-backed continuity	Project read/write access, native test/measurement tools, and an independent subagent or configured reviewer CLI for reviewed execution	npx skills add smota/tiny-skills --skill empirical-refinement-loop
software-engineering	git-deliver	Commit and push task-owned work, optionally integrate into the parent branch, and assess multi-agent worktrees	Git, repository access, commit identity, remote credentials for pushes, and project validation tools	npx skills add smota/tiny-skills --skill git-deliver
software-engineering	agent-refine	Distill session lessons and consolidate scoped agent-instruction layers with optional recursive and advisory-guidance review	Project read access and visible session context; write access only for confirmed changes	npx skills add smota/tiny-skills --skill agent-refine
software-engineering	project-init	Initialize project practices through a short intake, canonical instructions, ADRs, stack-aware validation, and optional multi-harness model tiering	Read/write access; selected toolchain for code; optional AFD and agent CLIs	npx skills add smota/tiny-skills --skill project-init
software-engineering	techlead-planner	Turn a goal into a reviewed plan before implementation: gate questions, vertical slices, ordering, architecture placement, and a non-functional pass	Project read access; optional read-only commands for evidence	npx skills add smota/tiny-skills --skill techlead-planner
software-engineering	cross-review	Run an independent, read-only adversarial review of a change in a different harness or model than the author's, and report findings with provenance	Git and repository read access; a second reviewer CLI (Claude Code, Codex, or pi) signed in; consent to send code to its provider	npx skills add smota/tiny-skills --skill cross-review
writing	release-notes	Write changelogs, release notes, and announcements for end-users, the technical team, or executives from selected commits, with flexible focus and presentation	None for supplied commit output; Git and repository read access for collection; optional hosted-source read access	npx skills add smota/tiny-skills --skill release-notes

Empirical refinement quick start

/empirical-refinement-loop run worker shutdown rounds=2

Runs up to two reviewed rounds on the scoped behavior, keeping evidence and stopping if blocked. See the skill documentation for init, plan, resume, prerequisites, and failure behavior.

Git delivery quick start

/git-deliver-merge develop

Delivers this task's commit to its remote branch, then integrates it into develop and pushes, proving each push on the remote. /git-deliver stops after the feature push and /git-deliver-audit inspects without changes. See the skill documentation for scope review, worktree ownership, and failure behavior.

Release communication quick start

/release-notes changelog since last release
For executives, brief, plain text.

Formats are changelog, notes, and announce; audiences are end-users, the technical team, and executives. The scope is a range, today, since last release, or supplied commit output, and every format uses the selected commits as its factual basis. See the skill documentation for evidence boundaries, prerequisites, and failure behavior. Generating text does not publish it.

Project initialization quick start

/project-init --plan .

Runs a read-only intake and proposes the baseline; --apply writes the reviewed choices and --verify checks an existing baseline. See the skill documentation for inputs, prerequisites, outputs, the optional multi-harness execution policy, and failure behavior.

Design

• Self-contained: skill instructions, references, examples, and command definitions live inside one skill directory.
• Independent: no shared runtime, prompts, state, or cross-skill assumptions.
• Standard: each skill exposes SKILL.md with standard frontmatter and can be consumed by compatible agent clients.
• Lean installs: a skill directory holds only what agents read or run; evals and refresh procedures live in maintenance/<skill-name>/.
• Simple: prefer small slash commands over frameworks or deployment infrastructure.
• Composable commands: use explicit command sections when one skill offers multiple related slash commands.

Skill structures are inspired by smota/metaskills --skill agent-builder, but intentionally constrained for straightforward maintenance and direct installation.

Repository layout

skills/
  personal/
    <skill-name>/
  research/
    <skill-name>/
  writing/
    <skill-name>/
  software-engineering/
    <skill-name>/

  # Each skill directory contains:
  # SKILL.md, optional README.md, references/, examples/, scripts/

templates/
  SKILL.md                # minimal starter template

docs/
  authoring.md            # authoring and validation rules

maintenance/
  <skill-name>/           # maintainer-only evals, refresh procedures, generators

Add skill

1. Choose a class: personal, research, writing, or software-engineering.
2. Copy templates/SKILL.md to skills/<class>/<skill-name>/SKILL.md.
3. Keep all context needed by the agent inside that directory.
4. Define simple slash commands under ## Commands; document arguments, output, and failure behavior. Ensure frontmatter description is quoted in "..." (or folded scalar >-), and set disable-model-invocation: true when only the user should start the skill.
5. Declare prerequisites and credits, and add examples for non-obvious workflows. Put evals and refresh procedures in maintenance/<skill-name>/.
6. Update the root README catalog with the class, purpose, prerequisites, and install command.
7. Validate locally with npx skills add . -l (confirming zero ⚠ Skipped warnings and that the skill appears in Available Skills), then test installation from the branch or commit with npx skills add.
8. Commit the skill and catalog update together.

Repository guidance

Repository structure and contribution rules live in AGENTS.md; authoring details live in docs/authoring.md.