Skip to content

About

Goal-QA-driven development toolkit for agentic AI

Resources

Stars

27 stars

Watchers

0 watching

Forks

Repository files navigation

agents-stack

A collection of composable agent skills, built for local-first execution and portable deployment.

Each directory is a standalone skill with its own SKILL.md protocol. Most skills can be copied into any repo independently. Directories stay at the repo root on purpose — grouping is by job, not by extra path depth.

Pick a skill

If you need Use
Fuzzy product / business / architecture decision meta-thinking-framework
What product to build (discovery, not implementation) interview-to-build
System prompt that keeps ignoring rules system-prompt-sop
UI / image / video generation without a design spec design-context
Pitch, cold email, landing copy, deck opener, value prop plain-sales-copy
Campaign / multimedia from a marketing-director brief marketing-director
Feature or bugfix through spec → code → evidence agentic-dev-e2e
Create or eval a new skill skill-creator
Graduate or deprecate a rule from real usage skill-evolution
Learn a domain to ~80% practitioner learn-anything-about-x
Local generate / read / memory agent-plugin/

Upstream of implementation: decide, then specify, then build. skill-creator owns short-cycle skill drafts; skill-evolution owns long-cycle case-driven change.

Decide

High-stakes thinking before anyone writes code.

Skill Purpose
meta-thinking-framework Six-lens analysis for complex, irreversible decisions
interview-to-build Map builder intuition onto a grounded discovery process

Specify

Turn a decision into a checkable contract.

Skill Purpose
system-prompt-sop Route each prompt constraint to its enforcement layer
design-context Design-context gate before frontend or visual generation
plain-sales-copy 3-5 sentence sales/business copy: no puff, four elements, "tell me more"
marketing-director Five-dimension brief, then multimedia with one job per asset

Build

Execute and review work in a repo.

Skill Purpose
agentic-dev-e2e End-to-end feature development and bug-fix workflow

Meta

Skills that create or evolve other skills and workflows.

Skill Purpose
skill-creator Create and evaluate agent skills (L0 generation, L1 evals)
skill-evolution Evolve skills from real usage evidence (L2 + L3)

Learn

Skill Purpose
learn-anything-about-x Adaptive domain mentor for structured learning

Agent Plugin

agent-plugin/ bundles eight runtime capabilities for local agents:

Skill Purpose
generate-image Generate or edit raster images via OpenRouter
generate-video Generate video jobs via OpenRouter
generate-text Generate text via OpenRouter
generate-audio Generate speech via OpenRouter
read-image Answer questions about local images
read-video Answer questions about local or remote video
read-audio Transcribe or analyze local audio
agent-memory Portable markdown memory: CoALA four drawers + optional persona core

Entry points:

  • scripts/openrouter.py -- provider adapter for all cloud-capable skills
  • scripts/doctor.py -- runtime prerequisite check

Storage:

  • Project-local default: ./.agents/
  • Global opt-in: ~/.agents/
  • Generated media, artifacts, and approvals live under the selected root.
  • Agent memory uses markdown drawers under .agents/memory/; no database is required.
  • Boot loads an optional persona plus current working memory; semantic / episodic / procedural load on demand.
  • Persona files are optional and must be user-confirmed.
  • ~/.agents/skills/ is reserved for installed skills and is never overwritten by runtime state.

Installation

To install the full pack, copy agent-plugin/ into your project:

cp -R agent-plugin ./.agents/skills/

To install a single skill, copy just its directory:

cp -R agent-plugin/agent-memory ./.agents/skills/agent-memory

To install a protocol skill from the repo root:

cp -R meta-thinking-framework ./.agents/skills/meta-thinking-framework

To install the memory drawer templates:

cp -R agent-plugin/templates/memory ./.agents/memory

Configure the provider:

cp agent-plugin/templates/openrouter.config.json .agents/openrouter.config.json
export OPENROUTER_API_KEY="your-key"

Runtime

The OpenRouter adapter is zero-dependency Python 3. Run from the pack root:

python3 scripts/doctor.py
python3 scripts/openrouter.py --help

Tests:

pytest agent-plugin/tests

Philosophy

Every skill is an independent capability. Shared contracts (provider-contract, storage-schema, privacy-policy, output-contract) keep them composable without coupling. Local-first means the adapter refuses cloud calls without explicit approval, and all generated state stays on disk in a predictable layout.

About

Goal-QA-driven development toolkit for agentic AI

Resources

Stars

27 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages