_   ___  ___  _  _  ___  _  ___     ___  ___   ___   _    _  _  _  ___
  /_\ / __|| __|| \| ||_ _|| |/ __|   |_ _|/ _ \ / _ \ | |  | |/ /| ||_ _|
 / _ \\__ \| _| | .` | | | | | (__     | || (_) | (_) || |__| ' < | | | |
/_/ \_\___/|___||_|\_||___||_|\___|    |_| \___/ \___/ |____|_|\_\|_| |_|

10 specialized agents · 12 commands & skills · Claude · Opencode · Ampcode · Droid — plus a full Linux terminal dev environment

The agent kits here started as scaffolding for early LLMs that needed to be told everything. They aren't that anymore. Models got smarter, so the toolkit got thinner: specialists you call by name or that trigger themselves on domain, wide lanes instead of tight rails, constraints on what rather than scripts for how.

They're actively challenged, pruned and rewritten — things get renamed, added, and deleted as models improve. That churn is the point.

Around them sits the rest of the box: per-distro Linux dev-environment setup, BYOK/model-switching customization, and a curated marketplace.

# Agents — interactive installer, auto-updates
npx liteagents

# or copy manually for your tool
cp -rv ai/subagentic/claude/*   ~/.claude/           # Claude Code
cp -rv ai/subagentic/opencode/* ~/.config/opencode/  # OpenCode
cp -rv ai/subagentic/droid/*    ~/.factory/          # Droid
cp -rv ai/subagentic/ampcode/*  ~/.config/amp/       # Amp

---

The catalog

Agents — invoke with @name (Claude / OpenCode / Amp) or invoke droid name.

Agent	What it's for
1-create-prd	Define scope as a PRD — a portal into the work, not a spec to obey
2-generate-tasks	Break a PRD into granular, actionable tasks
3-process-task-list	Execute tasks one at a time with review checkpoints
orchestrator	Read intent, route to the right agent sequence
code-developer	Implementation, debugging, refactoring
quality-assurance	Test architecture, quality gates, risk assessment
feature-planner	Epics, user stories, prioritization, backlog
market-researcher	Market and competitive analysis, project discovery
system-architect	System design, tech selection, API design, scale
ui-designer	UI/UX, wireframes, prototypes, design systems

Commands & skills — /name. Three of them can fire on their own when the situation matches: /brainstorming, /root-cause, /live-canvas.

Command	What it's for
/stash	Snapshot this session's context before compaction or handoff
/remember	Fold stashes + friction into hot project memory
/docs-builder	Reorg, index, and split a docs corpus so search actually finds things
/self-review	Verify what you delivered since the last self-review with real runs, and review its structure
/branch-review	Full pre-merge review, docs sweep — blockers reported, nits to the fix ledger
/release	CHANGELOG, version bump, local commit, then hand back the merge sequence
/refactor	Work the fix ledger's nit bullets; with args, refactor and optimize a named area
/security	Standalone vulnerability audit (also stage 2 of /branch-review)
/test-generate	Generate a test suite and verify each test exercises real code
/brainstorming	Turn a rough idea into a formed design by questioning
/root-cause	Find the cause before changing code — evidence, backward trace, one hypothesis, fix at the source
/live-canvas	UI variations with click-to-annotate feedback in the browser

_{Claude Code and Amp ship all 12 as skills; Opencode and Droid expose all 12 as commands. All four also ship agent reference docs.}

---

The ones that carry the load

AGENT_RULES.md — the architect behind every project

One lightweight, generic, model-agnostic rules doc. It sets what matters and leaves wide room on how: simple over clever, every line has a purpose, surgical changes, exhaust the stdlib before reaching for a dependency, POC before you design.

With smarter models a PRD is a portal, not a deliverable — the start of a conversation. You discover it module by module, reviewing, changing and deleting as experience arrives, instead of pinning everything down up front.

It is loaded every session, linked from CLAUDE.md, and /remember keeps it fresh — each run byte-compares your copy against the template that shipped with your install.

your copy	what happens
identical	nothing — no write, no output
missing	copied in
different	moved to AGENT_RULES.md.bak, template copied in, both reported

Edits are never destroyed, but they aren't preserved in place either. .bak is a single file the next update overwrites — a customised body survives one release, not two.

/stash → /remember — the daily driver

/stash  ──►  /remember
capture      consolidate

/stash is project-local: one session, one feature or scope or goal. Start fresh on delivery, or when context hits ~30% (never past 40%). The stash is the clean handoff to the next session.

/remember fires on the nudge every 5 stashes. It folds those 5 in — and it also reads across all your projects, mining session logs for where you and the model actually collided: corrections, dead ends, abandoned flows. Those become rules, so the next session is less wrong than this one.

Markdown files, no database, no RAG, always hot in context, and it keeps the nuances of each project separate. Runs on a mid-tier model.

/docs-builder — the other daily driver

Markdown gets out of hand. A token-hungry agent then can't find the knowledge that matters. (Its own PRD grew from 500 lines to 3k+ as findings piled in — exactly the problem it exists to solve.)

It indexes everything under /docs in four buckets — product/, logs/, wiki/, archive/ — reorganizes, indexes, and splits oversized docs so you don't have to. Mid-tier model.

/live-canvas — the UI every agent CLI was missing

A terminal agent can't see your screen, and you can't describe "that padding, on that card, but only on mobile" in words without burning ten minutes.

/live-canvas spins up one lightweight localhost page with your UI on it. You click the thing that needs changing, type what you want, and submit. Comment on as many elements as you like, all in one pass, then send the batch to the agent.

Two ways it comes back:

• JSON mode — feedback lands in a file, you tell the agent to read it. Works in any tool.
• Live mode — an MCP channel streams each comment straight into the session, so edits land while you're still in the browser. Claude Code only.

It also ships with real UI direction baked in, so it can generate variations of a screen for you to pick from — no more hours spent nudging divs to find out what you actually wanted.

/self-review → /branch-review → /release → /refactor

• /self-review — committed work since the last self-review (or exactly the commit hashes or a..b range you give it; a dirty tree stops), before /branch-review. The orchestrator only writes a handoff; one spawned mid-tier worker tries to break the claims with real runs (works, no regression, cleanup — dead code, state ownership, reuse, naming, performance — glossed over, underspecced) and reports max 5 failure-sentence items plus max 5 Cleanup items in Fix now / Later. It never fixes anything — whatever you don't fix now goes to the fix ledger, tagged nit, change or idea (missing, worth building). Not a gate.
• /branch-review — the powerhouse. Reviews every change on a branch, medium depth by default, or exactly the commit hashes or range you give it (no record, no docs sweep then). Surfaces confirmed blockers only: real bugs, test quality, plus a full OWASP-shaped security pass (no leaked keys, no injection, trust boundaries checked) that runs at full depth regardless of level. Everything non-blocking goes to the fix ledger. It also sweeps and commits the project's docs — README, PRD, findings — for what the branch changed, once at the end, once the review is settled (ready, or every blocker pushed through by name).
• /release — does the last pre-release chore: runs short mechanical checks (lint, migrations, in sync with origin; tests only if the review record's tests: line doesn't cover them), writes the CHANGELOG entry, bumps the version, commits locally. Then it tells you you're ready to merge, and hands the sequence back. It never pushes.
• /refactor — with no arguments, works the fix ledger: fixes nit bullets, lists change and idea bullets (bigger than a refactor) and asks you to keep, drop or spec each. Cumulative by design: nits pile up until you choose to clear them, so review and release never drown in them. With a whole-area argument it first lists candidates (what is wrong, the proposed change, strength) and stops; only the ones you pick are edited.

---

The rest of the box

Linux dev environment

./tools-fedora/dev_tools_menu.sh    # Fedora
./tools-debian/dev_tools_menu.sh    # Debian / Ubuntu

An interactive menu that installs and configures a terminal dev setup — tmux, Neovim/LazyVim, zsh, fzf, lazygit, Kitty/Ghostty. Per-tool guides live next to the scripts.

Customization (ai/customize/)

BYOK keys, a Claude Code LLM/MCP switcher, Ollama configs, and the shared agent guidelines.

Marketplace (ai/marketplace/)

Curated subagents, plugins, skills, MCP servers, and workflows.

Layout

ai/
  subagentic/   # agent kits: claude, opencode, droid, ampcode (+ manual)
  customize/    # byok, claude-switcher, ollama, config, skill-to-command
  marketplace/  # curated subagents, plugins, skills, mcp, workflows
tools-debian/   # Linux dev-tool setup (Debian / Ubuntu)
tools-fedora/   # Linux dev-tool setup (Fedora)
docs/           # guides

---

Deeper docs

Document	What's in it
/remember	The /stash → /remember pipeline, friction sensor, antigen ledger
/docs-builder	Reorg and cleanup modes, measured cost, the drift ledger
/self-review	The handoff → worker → relay flow, the bar, Fix now / Later
/branch-review	The four stages, what blocks, the fix-ledger loop
/live-canvas	Both modes, the click-to-annotate overlay, and setup
live-canvas-channel	The Claude Code MCP channel plugin — install, protocol, debugging
AGENT_RULES.md	The rules doc itself
All agents & commands	Full reference, token loads, progressive disclosure
Vibecoding 101	Beginner's guide to AI-powered development

---

Apache-2.0 © 2026 hamr0 — see LICENSE.