Skip to content
/

Ozzeron prompt-pack

ozzeron-prompt-pack · Ozzeron/prompt-pack · ★ 6 · last commit 2026-05-13

Primitive shape 23 total
Skills 23
00

Summary

Ozzeron/prompt-pack — Summary

prompt-pack is an opinionated, curated discipline pack of 23 AI coding-agent skills focused on eliminating technical debt patterns AI agents generate by default: duplicate components, convention drift, scope creep, and "helpful" rewrites of working code. The pack's central philosophy is "reuse before create" — every code-creating role inherits a reuse-before-create meta-skill that mandates searching for existing artifacts before adding new ones, with a mandatory one-line justification for any new artifact created. Unlike sprawling awesome-cursorrules collections (hundreds of unaudited entries), prompt-pack ships 23 curated, lint-gated, cross-tool skills that have been reviewed across nine external review rounds plus four empirical field tests on real codebases. It distributes via shell/PowerShell installers with nine targets (Cursor 2.4+ Skills, universal .agents/skills/, Claude Code, Codex, OpenClaw, and more) and six profiles — making it tool-agnostic. The task-router skill provides orchestration for legacy/orchestrated flows, mapping user intents to specific roles including composed flows (PR review = code-review → security-review).

differs_from_seeds: Closest to superpowers (skills-only behavioral framework with no commands), but prompt-pack inverts the philosophy — where superpowers uses Iron Law enforcement, prompt-pack focuses specifically on anti-debt discipline (reuse-before-create, convention discovery, token-disciplined reads) rather than comprehensive workflow coverage. Unlike spec-kit's mirror command+skill pattern, prompt-pack has no commands at all. Unlike BMAD-METHOD's persona model, prompt-pack's skills are role-agnostic discipline rules, not agent personas. The distinctive engineering choice is the mandatory justification trail and inheritance hierarchy (meta skills inherited by all coding roles) which no seed framework implements.

01

Overview

Ozzeron/prompt-pack — Overview

Origin

Created by GitHub user Ozzeron as a response to the "catalogue anti-pattern" in existing AI agent skill collections. Published on GitHub under MIT license, first public commit circa 2025, actively maintained with current version 0.4.1 (as of May 2026, 6 stars).

Philosophy

The pack's core thesis, quoted verbatim from the README:

"The usual collections (awesome-cursorrules, awesome-claude-code-subagents, awesome-copilot, awesome-agent-skills, etc.) are catalogues: hundreds or thousands of prompts, every variant of 'you are a senior X', optimised for breadth and discoverability. Pick one, paste it in, hope."

"prompt-pack picks the opposite trade-off: fewer skills, stronger behaviour. 23 curated skills, lint-gated, native skill discovery..."

"Under those prompts, agents ship the same recurring failures regardless of model: duplicate components, fresh utilities for things that already exist in the repo, dependencies added for problems the project already solved, convention drift, scope creep, and 'helpful' rewrites of code that was fine. Bigger catalogues do not fix this; they multiply the surface where it can happen."

The Five Disciplines

  1. Reuse before create — mandatory search before any new artifact; one-line justification required
  2. Convention discovery first — architecture skills inspect 2-3 canonical examples before writing code
  3. Attention-disciplined reads — explicit scope limits + "don't read these" rules per skill
  4. Orchestrator-aware routingtask-router maps intent to roles including composed flows
  5. Curated, not exhaustive — every skill must earn its place

Anti-patterns Explicitly Named

  • Creating new artifacts when existing ones would do
  • Reading everything in context (context = attention budget, not just token count)
  • Treating breadth of catalogue as a quality signal
  • "Helpful" rewrites of working code

Language Primary

PowerShell (installer), Node.js (linter), Markdown (skills)

02

Architecture

Ozzeron/prompt-pack — Architecture

Distribution

  • Type: skill-pack (standalone repo + shell installers)
  • Language: PowerShell (Windows), Bash (Linux/macOS), Node.js (linter)
  • Version: 0.4.1
  • License: MIT

Install

git clone https://github.com/Ozzeron/prompt-pack.git ~/code/prompt-pack
cd ~/code/your-project
~/code/prompt-pack/install.sh --target cursor --profile minimal

PowerShell equivalent:

git clone https://github.com/Ozzeron/prompt-pack.git $HOME\code\prompt-pack

Nine install targets: cursor (2.4+ native Skills), cursor-foundation (foundation only), cursor-legacy (rules), agents (universal .agents/skills/), claude-code, codex-skills, codex-legacy (AGENTS.md), openclaw, paste (raw copy)

Six profiles: minimal, backend, frontend, fullstack, review, all

Directory Structure

prompts/
  architecture/      # backend-api, frontend-feature, database-schema,
                     # database-migrations, postgres-supabase, refactor-planner
  review/            # code-review, repo-audit, security-review,
                     # frontend-audit, database-review,
                     # duplication-audit, debugger
  interface/         # ui-designer
  delivery/          # handoff, test-writer, doc-writer, ai-agent-docs
  infra/             # docker
  meta/              # task-router, engineering-principles, reuse-before-create,
                     #  token-discipline
docs/
  USAGE.md
  CONTRIBUTING.md
  PROMPT-FORMAT.md
scripts/
  lint-skills.mjs    # format linter run on every PR
install.sh
install.ps1

Each prompt is a directory:

prompts/<category>/<name>/
  SKILL.md           # prompt + YAML frontmatter
  EXAMPLES.md        # optional
  CHANGELOG.md       # optional

Required Runtime

  • node >= 18 (for linter only — not required for consuming skills)
  • No runtime required to use skills; agents read SKILL.md files directly

Target AI Tools

  • Cursor 2.4+ (native skill discovery via description matching)
  • Claude Code (subagent mode)
  • Codex CLI
  • OpenClaw
  • GitHub Copilot (via .agents/skills/)
  • Generic / copy-paste
03

Components

Ozzeron/prompt-pack — Components

Skills (23 total)

Meta category (4)

Name Purpose
task-router Maps user intents to specific roles; composes flows (e.g. PR review = code-review → security-review)
engineering-principles Core DRY/quality rubric inherited by all review roles
reuse-before-create Central anti-debt discipline; mandates search before any new artifact creation
token-discipline Context-budget discipline; diff-first, full-files only when needed

Architecture category (6)

Name Purpose
backend-api API layer design using convention discovery
frontend-feature UI feature scaffolding with canonical-example inspection
database-schema Schema design with project convention inspection
database-migrations Migration writing with safety checks
postgres-supabase Postgres/Supabase-specific patterns
refactor-planner Safe refactor planning with risk assessment

Review category (7)

Name Purpose
code-review Severity-classified PR review; correctness + security + maintainability + tests
repo-audit Full codebase audit (not diff-based)
security-review Security-focused review pass
frontend-audit UI/UX code audit
database-review Database schema and query review
duplication-audit Cross-file duplication detection
debugger Systematic debugging with root-cause isolation

Interface category (1)

Name Purpose
ui-designer UI design and component specification

Delivery category (4)

Name Purpose
handoff Session handoff documentation generation
test-writer Test suite authoring
doc-writer Documentation authoring
ai-agent-docs AI-agent-readable docs (AGENTS.md/CLAUDE.md authoring)

Infra category (1)

Name Purpose
docker Docker configuration and Dockerfile authoring

Commands

None — this is a skills-only pack. Discovery is via host skill system (Cursor) or task-router.

Hooks

None — no Claude Code lifecycle hooks.

Scripts

Name Purpose
scripts/lint-skills.mjs CI linter; validates YAML frontmatter format on every skill; runs on every PR
install.sh Multi-target, multi-profile installer for Linux/macOS
install.ps1 Multi-target, multi-profile installer for Windows/PowerShell

Templates

None shipped separately (SKILL.md format defined in docs/PROMPT-FORMAT.md)

Subagents

None — no subagent spawn mechanism. task-router composes skills sequentially but does not spawn agents.

05

Prompts

Ozzeron/prompt-pack — Prompt Excerpts

Excerpt 1: reuse-before-create (meta/reuse-before-create/SKILL.md)

Technique: Strict ordered preference list with non-negotiable enforcement + mandatory justification trail

---
name: reuse-before-create
description: DRY discipline core. Before creating anything new, find what already exists. Every coding skill inherits this.
category: meta
version: 0.1.0
triggers: ["writing code", "adding component", "adding util", "creating file", "new helper", "new hook"]
applies_to: [openclaw, cursor, claude-code, generic]
---

# Reuse Before Create

## The rule

**Before writing anything new, search for what already exists. Reuse first; create last.**

The order of preference is strict and non-negotiable:

1. **Reuse as-is** — existing artifact fits, import and use
2. **Extend** — add a prop, option, generic, or branch to existing artifact
3. **Compose** — combine existing artifacts into a new shape without rewriting them
4. **Extract** — pull a shared core out of two near-duplicates; both use the new core
5. **Create new** — only when none of the above apply

Each step further down costs more — both in code volume and in long-term maintenance.
Skipping straight to "create new" without proving the prior steps don't apply is the
default failure mode this skill prevents.

## Justification is not optional

When you do create something new, **a one-line justification is mandatory**, not a
nice-to-have. It belongs in the commit message, the PR description, the handoff, or as
an inline comment near the new artifact.

If you can't write the justification in one sentence, you haven't searched hard enough.
Go back to step 1.

Analysis: Uses ordered enumeration with explicit "non-negotiable" language; enforces justification as a required output artifact; prevents the agent from silently skipping to creation. The triggers frontmatter field enables host-native skill discovery in Cursor without a router.


Excerpt 2: code-review (review/code-review/SKILL.md)

Technique: Preflight checklist pattern — mandatory steps before reading any code; prevents hallucinated "reviews" of files the agent hasn't properly read

## Preflight (do this before reading any source file)

This is a checklist on purpose. Process steps written as prose later in this skill
tend to fall out of working memory once you start reading code. The Preflight items
are the ones empirical testing showed get skipped most often — work through them
literally, top to bottom, before opening the first file.

- [ ] **Routing check.** Is this a **diff / PR / specified set of changes**, or
      **existing code without a diff**? If the user did not provide a diff and is asking
      you to look at existing files, this is an audit, not a review — stop and route
      to `review/frontend-audit` (or another `*-audit` skill), or ask one clarifying
      question.
- [ ] **Reading plan.** Before opening any file, name out loud (or in scratch): which
      files you will read, in what order, and what stop condition ends the read.
- [ ] **Project conventions, once.** Read `AGENTS.md` / `CLAUDE.md` / project `README`
      once at the start.
- [ ] **All four passes are mandatory.** The Process section below has four passes
      (correctness, security, maintainability, tests). Pass 4 (tests) is the one most
      often dropped silently. If the diff has no tests, you say so explicitly.

If you cannot tick all four, stop and ask. Do not start reading.

Analysis: Checkbox-based preflight borrowed from aviation/surgery checklists; accounts for working memory limitations; calls out empirically-observed failure mode (pass 4 dropout); uses self-routing instruction to send audit requests to the correct skill.


Excerpt 3: Inheritance Declaration

Both skills above include an ## Inherits section:

## Inherits

- [`meta/engineering-principles`](../../meta/engineering-principles/SKILL.md)
- [`meta/token-discipline`](../../meta/token-discipline/SKILL.md)

Technique: Explicit skill inheritance graph — code-review imports both engineering-principles (for the rubric) and token-discipline (for read-order rules). This is not executed at runtime — it is a documentation contract the agent must honor when reading the skill.

09

Uniqueness

Ozzeron/prompt-pack — Uniqueness & Positioning

differs_from_seeds

Closest to superpowers (both are skills-only, no commands, no hooks), but prompt-pack's editorial philosophy inverts superpowers' comprehensiveness. Superpowers covers the entire SDLC with 14 skills including spec-writing, workflow, and verification. prompt-pack focuses specifically on the debt-prevention niche: its central innovation is the reuse-before-create inheritance chain, where every code-creating skill carries an anti-duplication precondition. Unlike spec-kit's mirror command+skill architecture, prompt-pack has no commands at all. Unlike BMAD-METHOD's persona model (you are a senior architect), prompt-pack skills are behavioral constraints applied during any agent session, not role assignments. The nine install targets and six profiles represent a cross-tool portability story none of the seeds attempt — openspec supports multiple tools but through separate config files, not a unified installer.

Positioning

  • Anti-catalogue: Explicitly positioned against awesome-cursorrules and similar large collections
  • Debt-prevention over feature-coverage: 23 skills vs. hundreds; each earns its place
  • Attention discipline: Frames context management as signal-per-token, not token-minimization
  • Cross-tool first: Nine targets as a design constraint, not an afterthought

Observable Failure Modes

  1. Justification-only enforcement: The reuse-before-create justification requirement is enforced by prompt text, not tooling — agents that compress or skip skill content will miss it
  2. No state: No cross-session continuity means convention-discovery insights from one session don't persist to the next
  3. No verification loop: Skills produce outputs but do not automatically validate them (no test-runner hook, no linter invocation)
  4. Breadth gaps: 23 skills covers common patterns but leaves infra-heavy stacks (k8s, Terraform, complex CI) unaddressed — by design, but a real gap for some teams

Most Distinctive Opinion

Context is an attention budget. A model that reads everything is a model that mixes patterns from unrelated code and hallucinates with confidence. Token discipline and scope limits are correctness improvements, not just cost optimizations.

04

Workflow

Ozzeron/prompt-pack — Workflow

Workflow Pattern

prompt-pack does not define a linear project workflow. It provides a library of discipline-enforcing skills that are invoked individually or via task-router. The closest thing to a workflow is:

1. Install pack (choose target + profile)
2. Start coding task in AI agent
3. Agent activates relevant skill via description matching (Cursor native) or task-router (legacy)
4. Skill enforces reuse-before-create, convention discovery, and token discipline
5. Agent produces code / review / artifact
6. (Optional) PR review = code-review → security-review composed flow via task-router

Phases

Phase Artifact
Setup Skills installed to target path (.agents/skills/, .cursor/rules/, etc.)
Architecture backend-api, frontend-feature, database-schema outputs
Review code-review, security-review, duplication-audit outputs
Delivery handoff notes, test files, documentation

Approval Gates

None — no formal approval gates defined. The reuse-before-create skill creates a soft gate: the agent must produce a justification before creating any new artifact, but this is not enforced by tooling — it is enforced by the skill prompt text.

Spec Format

None — prompt-pack does not generate specs. Skills are invoked during implementation, not during planning.

Composed Flow Example (task-router)

PR review composition:

  1. task-router identifies intent as "review this PR"
  2. Routes to review/code-review → produces severity-classified findings
  3. Routes to review/security-review → adds security pass findings
  4. Agent returns merged report
06

Memory Context

Ozzeron/prompt-pack — Memory & Context

State Storage

None — prompt-pack does not define any persistent state mechanism. It is a stateless skill library.

Memory Persistence

None. Each skill invocation is independent. The handoff delivery skill generates a handoff document, but this is a human-readable artifact, not a machine-parseable state file.

Context Management Philosophy

The pack has an explicit and documented stance on context management — unique in that it frames context as an attention budget rather than a cost center:

"Attention-disciplined, not token-cheap. Every prompt has explicit scope limits and 'don't read these things' rules, because context is an attention budget and a window stuffed with low-signal files makes the model worse — even at unlimited cost. Bigger context is not smarter context."

This is operationalized via meta/token-discipline:

  • Read diff first, not full files
  • Identify which files reduce uncertainty before opening them
  • Explicit "don't read" rules per skill (e.g. code-review says "Read the diff first, not the full files")
  • Full-file read only when correctness demands it

Cross-Session Handoff

The delivery/handoff skill generates a handoff document when invoked. Contents are human-defined per project; no standard schema is enforced.

Compaction Handling

No compaction mechanism. Skills are short enough to be read in a single context window.

07

Orchestration

Ozzeron/prompt-pack — Orchestration

Multi-Agent

No — prompt-pack is a single-agent skill library. No subagent spawning.

Orchestration Pattern

Sequential (via task-router). The task-router skill maps user intent to one or more roles and can compose sequential flows (e.g., code-review → security-review), but this is single-agent sequential execution — not multi-agent parallelism.

Isolation Mechanism

None. Skills operate in-place in the agent's current session context.

Multi-Model

No. The pack is model-agnostic and does not define model routing. It works with any AI coding tool.

Execution Mode

One-shot (each skill invocation is independent). The task-router composes multi-step flows within a single session, but there is no persistent loop.

Consensus Mechanism

None.

Prompt Chaining

Limited: task-router chains code-review → security-review output, but the intermediate output is not formally structured as input to the next stage — it is handled sequentially by the same agent session.

Cross-Tool Portability

High — nine install targets, six profiles. Works with Cursor, Claude Code, Codex, OpenClaw, Copilot, generic. The SKILL.md format with YAML frontmatter and applies_to field is the portability mechanism.

08

Ui Cli Surface

Ozzeron/prompt-pack — UI & CLI Surface

CLI Binary

No dedicated binary. The package.json has no bin field — it is a private package ("private": true). Install is via install.sh / install.ps1 scripts invoked manually.

Local UI / Dashboard

None.

IDE Integration

Works via native IDE skill discovery (Cursor 2.4+ uses description frontmatter field for matching) or via task-router for legacy flows. No VSCode extension or IDE fork.

Linter / CI Tooling

The pack ships a Node.js skill linter (scripts/lint-skills.mjs) that runs on every PR via GitHub Actions. It validates YAML frontmatter format and enforces the SKILL.md schema. This is a developer-facing quality gate, not an end-user tool.

Observability

None. No audit log, no session traces, no cost tracking.

Installation UX

The shell/PowerShell installers accept --target and --profile flags:

  • --target: one of nine deployment targets
  • --profile: one of six curated profiles (minimal, backend, frontend, fullstack, review, all)

This is the primary UX surface — choosing where skills land and which subset to install.

Related frameworks

same archetype · same primary tool · same memory type

claude-mem (thedotmack) ★ 78k

Background worker service captures every tool call as an observation, AI-compresses sessions, and auto-injects relevant past…

pi (badlogic/earendil) ★ 55k

A minimal, hackable, multi-provider terminal coding agent that adapts to your workflows via npm-installable TypeScript Extensions…

Agent Skills (Addy Osmani) ★ 46k

Encodes senior-engineer software development lifecycle as 23 auto-routed skills and 7 slash commands for any AI coding agent.

wshobson/agents Plugin Marketplace ★ 36k

Single Markdown source for 83 domain-specialized plugins that auto-generates idiomatic artifacts for five AI coding harnesses.

TabbyML/Tabby ★ 34k

Self-hosted AI coding assistant server (alternative to GitHub Copilot) with admin dashboard, RAG-based completions, and multi-IDE…

Compound Engineering ★ 17k

Make each unit of engineering work compound into easier future work via brainstorm→plan→execute→review→learn cycles.