A plugin marketplace for Claude Code and Cowork with 16 specialized AI agents for product management, business analysis, concept validation, interview coaching, and strategic thinking. Every agent writes its full analysis to ./outputs/ as a standalone deliverable.
/critic and /debate are the highest-value agents in the kit. Use them early and often.
/critic is specifically prompted for anti-sycophancy. Most AI tools agree with whatever you say. This agent does the opposite: it pushes back on your reasoning, won't soften feedback to preserve rapport, and won't tell you your idea is strong unless it actually is. Two modes: coaching (exposes blind spots, forces you to defend your thinking) and document review (6-pass adversarial analysis with a Ship/Fix/Rethink verdict). Callable at any point in the pipeline.
/debate assembles user-scoped expert panels. You choose the domain (business/venture, technical, specialty technical, financial, customer/market, or mixed), and the agent asks clarifying questions before selecting experts. Supports parallel panels: run a VC panel and a technical panel simultaneously, then get a cross-panel synthesis showing where they agree, disagree, and what each missed.
Tree of Thought (ToT): Several agents use Tree of Thought, a prompting technique where the AI generates multiple distinct reasoning paths in parallel, evaluates each one, cross-pollinates the strongest ideas, and converges on a synthesized answer. Instead of one linear response, you get the equivalent of multiple experts debating and refining. Agents using ToT: /consult (3 strategic approaches scored on desirability/viability/feasibility), /strategy (3 market entry strategies with branching outcomes scored on profitability/scalability/risk), /debate (expert panel with iterative challenge rounds).
Socratic Coaching: /bizmodel, /pricing, /survey, and /personas use Socratic method — they coach through questioning rather than handing you answers. The founder discovers their own insights through the framework.
Anti-Sycophancy Prompting: /critic is explicitly prompted to counteract AI agreeableness. It will not soften feedback, perform enthusiasm it doesn't hold, or lead with praise as a cushion before criticism.
Add the marketplace and install the plugin:
/plugin marketplace add jrenaldi79/plugin-marketplace
Then browse and install:
/plugin
There is a known bug where third-party GitHub marketplace plugins get wiped on restart for personal (non-org) Cowork accounts. The install script below works around this by writing directly to the local persistence path that survives restarts. It also sets up a scheduled task that checks for updates daily at 9am.
Prerequisites: Git and Python 3.10+ must be installed and on your PATH.
Open Terminal and run:
git clone https://github.com/jrenaldi79/plugin-marketplace.git && python3 plugin-marketplace/scripts/install-product-kit.py
Open PowerShell and run:
git clone https://github.com/jrenaldi79/plugin-marketplace.git
python plugin-marketplace\scripts\install-product-kit.py
After the script finishes, restart Claude Desktop. The plugin and scheduled task will both appear.
Optional: Add the marketplace to the Browse UI. The install script makes the plugin functional, but to see the marketplace in Customize > Browse, you also need to add it through the UI once: open Cowork, go to Customize > Browse Marketplace > Add Marketplace, and enter jrenaldi79/plugin-marketplace. This is a one-time step (the server remembers it).
- Finds your Cowork data directory (auto-discovers account/org paths)
- Clones the marketplace repo
- Extracts plugin files into the local cache
- Registers the plugin in the persistence files that survive restarts
- Creates a scheduled task (
update-product-kit-auto) that runs daily at 9am to check for updates
If the scheduled task hasn't run yet or you want to update immediately, pull the latest and re-run the script. The repo is already cloned from your first install.
macOS / Linux:
cd plugin-marketplace && git pull && python3 scripts/install-product-kit.py
Windows PowerShell:
cd plugin-marketplace
git pull
python scripts\install-product-kit.py
Check for updates without installing:
python3 plugin-marketplace/scripts/install-product-kit.py --check
Install without the scheduled task:
python3 plugin-marketplace/scripts/install-product-kit.py --no-schedule
&& not recognized (Windows PowerShell)
Older versions of PowerShell don't support &&. Run the two commands separately:
git clone https://github.com/jrenaldi79/plugin-marketplace.git
python plugin-marketplace\scripts\install-product-kit.py
python not found or Fatal Python error: Failed to import encodings module
This usually means Python isn't installed, isn't on your PATH, or has a broken installation (common with Python 3.14 on Windows). Fix:
- Install Python 3.12 from python.org/downloads
- During install, check "Add python.exe to PATH"
- Close and reopen PowerShell
- If an older broken Python is still taking priority, find the correct path with
Get-Command python -Alland run it directly, for example:C:\Python312\python.exe plugin-marketplace\scripts\install-product-kit.py
git not found
Install Git from git-scm.com. On macOS, you can also run xcode-select --install.
Plugin shows old version after update Run the install script again. It will fetch the latest from GitHub and update the cache. Restart Claude Desktop after.
Restarting on Windows Closing the Claude window with the X button does not quit the app. It keeps running in the system tray. To fully restart: right-click the Claude icon in the system tray (bottom-right of your taskbar, you may need to click the ^ arrow to find it) and select Quit, then reopen Claude.
"Virtual Machine Platform" error (Windows) Cowork runs in a VM that requires virtualization support. If you see "Claude's workspace requires Virtual Machine Platform, but the virtualization service isn't responding," restart your computer (not just Claude). If the error persists after reboot, open PowerShell as Administrator and run:
Enable-WindowsOptionalFeature -Online -FeatureName VirtualMachinePlatform -NoRestart
Then restart your computer again.
Scheduled task doesn't appear The task only shows up after fully restarting Claude Desktop (see above for Windows). On the first run you'll need to approve tool permissions manually, then future runs are automatic. To ensure the task runs even when your computer is idle, open the task in Scheduled and turn on the "Keep awake" slider.
| Agent | Command | Description | Deliverable |
|---|---|---|---|
| YC Review | /yc-review |
YC-style office hours that pressure-test product concepts through six forcing questions: demand reality (is there pull?), status quo analysis (why is "do nothing" good enough?), desperate specificity (who is the most desperate user?), narrowest wedge (smallest thing you could charge for?), founder observation (what have you seen that others missed?), and future-fit (structural tailwind or point solution?). Produces a verdict, strongest element, biggest gap, and concrete next steps. Built on gstack office-hours framework. | outputs/yc-review-*.md |
| VC Review | /vc-review |
Investor-grade diligence combining gated screening (Elevator Clarity, Problem Severity, TAM, Timing Catalyst), deep analysis (Delta 4, Problem Decomposition, Solution-Problem Fit, Competitive Positioning, Pre-Mortem), and five BMAD adversarial stress tests (First Principles, Reverse Brainstorming, Six Thinking Hats, Red Team vs Blue Team, Analogous Company Analysis) capped by a Debate Club Showdown. Designed to run in parallel with /yc-review. Adapted from Jonathan Ellis / Sandalphon Capital's VC diligence pipeline. |
outputs/vc-review-*.md |
| CEO Review | /ceo-review |
Founder-mode plan review with four distinct modes: Scope Expansion (dream big — push toward the 10-star experience), Selective Expansion (hold scope but cherry-pick high-impact additions), Hold Scope (maximum rigor — trace every failure mode, edge case, and shadow path), and Scope Reduction (surgeon mode — cut to the minimum viable core). Enforces nine prime directives including zero silent failures, named error handling, and mandatory diagrams. Built on gstack plan-ceo-review framework. | outputs/ceo-review-*.md |
| Business Consultant (ToT) | /consult |
Three expert consultants and a skeptical risk analyst evaluate a business challenge through a five-phase Tree of Thought process: branch generation (3 distinct approaches), branch exploration (D/V/F assessment per approach), cross-branch evaluation and pruning, convergence on the optimal path, and deep-dive execution planning. Each approach is scored on desirability, viability, and feasibility. | outputs/consult-*.md |
| Market Researcher | /research |
Context-aware research agent with three modes: Market Research (TAM/SAM/SOM sizing with top-down and bottom-up triangulation, landscape mapping, trend analysis with confidence levels), Competitive Intelligence (feature/pricing teardowns via web scraping, positioning gap analysis, funding/growth signals from Crunchbase and job postings), and Domain Research (industry terminology, regulatory landscape, standards, stakeholder maps, common failure modes). Scans uploaded docs and ./outputs/ first, scores evidence density across five dimensions, and only researches what's genuinely missing. Uses Tavily for search and Firecrawl for deep extraction when available; falls back to native WebSearch/WebFetch otherwise. Every claim is sourced and confidence-rated. |
outputs/research-*.md |
| Critic | /critic |
Dual-mode agent callable at any point in the pipeline. Specifically prompted for anti-sycophancy — will not soften feedback, perform enthusiasm, or lead with praise as a cushion. Coaching mode: brutally honest strategic coaching — extracts context and goals, exposes blind spots and faulty reasoning, emulates the top 0.01% domain expert (user picks who), builds a prioritized action plan, and highlights the single most impactful next step. Document review mode: point it at any file (especially a PRD) for a 6-pass BMAD-influenced adversarial analysis covering adversarial findings (minimum 10), edge case hunting, internal consistency checks, executability tests, hard questions, and a Ship/Fix/Rethink verdict with ranked findings. | outputs/critic-*.md |
| Business Model Architect | /bizmodel |
Socratic business model coaching grounded in three frameworks. Business Model Canvas: maps and diagnoses all 9 building blocks, surfaces misalignments (revenue stream vs. customer segment, cost structure vs. value proposition), identifies reinforcement loops. Ten Types of Innovation: pushes founders past product-only thinking to innovate across configuration (profit model, network, structure, process), offering (product system), and experience (service, channel, brand, engagement). 50+ Business Model Patterns: contextually introduces relevant patterns (multi-sided platform, freemium, product-as-a-service, pay-per-outcome, etc.) with real company analogies. Includes Blue Ocean Four Actions Framework (eliminate/reduce/raise/create) and a 5-point model stress test covering unit economics, scalability, defensibility, revenue model fit, and assumption stack. Socratic method — coaches through questioning, does not write the model for the founder. | outputs/bizmodel-*.md |
| Pricing Strategist | /pricing |
Socratic monetization coaching grounded in Monetizing Innovation (Ramanujam & Tacke). Flips the default sequence: price and market BEFORE you design and build. Diagnoses the four monetization failures — Feature Shocks (overengineered one-size-fits-all), Minivations (right product, priced too low), Hidden Gems (blockbuster feature buried or given away), and Undeads (no market pull at any price). Enforces the 9 Rules of Monetization: WTP validation (Van Westendorp, conjoint, purchase simulation), needs-based segmentation (reject demographics), Leader-Filler-Killer feature classification with Good-Better-Best configuration, monetization model selection (subscription, usage-based, freemium, dynamic, auction), pricing strategy declaration (maximization vs. penetration vs. skimming), outside-in business case (price elasticity, not cost-plus), value communication (benefits over features), behavioral pricing tactics (compromise effect, anchoring, decoy pricing), and price integrity (three non-pricing actions before any price cut). Cross-references /bizmodel Revenue Streams and /personas WTP signals. Includes a pricing architecture stress test and connects back to the full business model canvas. |
outputs/pricing-*.md |
| Expert Debate | /debate |
Assembles a user-scoped panel of domain experts to stress-test a problem. Asks the user to choose panel type first: Business/Venture (VCs, operators, strategists), Technical (engineers, architects), Specialty Technical (user names the field), Customer/Market (researchers, buyers, channel partners), Financial (CFOs, modelers, pricing experts), or Mixed/Custom. Supports parallel panels (e.g., run a VC panel and a technical panel simultaneously, then synthesize where they agree and disagree). Facilitates four debate rounds: initial perspectives, constructive challenges, assumption stress-testing, and convergent synthesis. Each expert speaks in their authentic voice. Cross-panel synthesis identifies green lights, risk zones, and blind spots. | outputs/debate-*.md |
| Agent | Command | Description | Deliverable |
|---|---|---|---|
| Interview Coach | /coach |
Scores and coaches customer discovery interviews against a research rubric grounded in three textbooks (Interviewing Users by Portigal, Jobs to Be Done, and The Mom Test concepts). Fetches reference materials at runtime, then evaluates the transcript on question quality, active listening, bias avoidance, and sentiment awareness. Provides a 1-10 score with category breakdowns and specific improvement recommendations. | outputs/coach-*.md |
| Interview Summary | /summarize |
Generates structured topline reports from interview transcripts with eight sections: executive summary, prioritized problems and pains (with emotional intensity markers), Jobs to Be Done analysis, key quotes, unspoken needs and underlying motivations, sentiment arc, strategic implications, and recommended next actions. Designed to produce research-ready artifacts a product team can act on immediately. | outputs/summary-*.md |
| Survey Design Coach | /survey |
Socratic process for designing closed-ended quantitative surveys. Follows a strict four-step methodology: diagnose context (product concept, target audience, assumptions), explore quantitative methods (scaling, ranking, MaxDiff, budget allocation), collaboratively draft questions measuring problem resonance, problem prioritization, and willingness to invest, then finalize with a ready-to-deploy survey instrument. Will not suggest question wording until context is fully diagnosed. | outputs/survey-*.md |
| Persona Developer | /personas |
Four-phase guided process that transforms business data into detailed buyer personas. Phase 1: business and market assessment (5 structured inputs). Phase 2: segment identification (2-4 segments scored on LTV, acquisition difficulty, and size). Phase 3: deep persona development (core identity, psychological profile, problem/solution context, buying journey, daily experience). Phase 4: strategic implementation guide (messaging, channels, product insights, sales enablement). | outputs/personas-*.md |
| Agent | Command | Description | Deliverable |
|---|---|---|---|
| PRD Builder | /prd |
Context-harvesting PRD generator. Scans ./outputs/ for prior agent deliverables (yc-review, consult, personas, strategy, etc.) and pre-fills sections automatically — only prompts the user for genuinely missing information. Covers 11 sections: problem & vision, goals, constraints/assumptions/out-of-scope, user personas, narrative user journeys (story-based with opening/trigger/core experience/edge cases/resolution), functional requirements (FR-numbered, 15-40 target), non-functional requirements (performance, security, accessibility, reliability), success metrics with targets, technical considerations, milestones & sequencing with go/no-go criteria, and user stories with sprint-readiness flags. Runs a self-validation pass (completeness, quality, traceability) before finalizing. |
outputs/prd-*.md |
| Meta-Prompt Engineer | /prompter |
Collaboratively designs and refines AI system prompts through a structured process: deconstruct the goal (clarifying questions about function, audience, inputs/outputs), identify key components (identity, objective, ruleset, workflow), draft the prompt with deliberate persona and tool integration, then iteratively refine based on testing and feedback. Follows principles of clarity over brevity, structure over prose, and tool-agnostic design. | outputs/prompter-*.md |
| Skill | Description |
|---|---|
using-product-kit |
Index skill — helps Claude route to the right agent based on natural language requests |
Every agent saves its complete analysis to ./outputs/ as a standalone markdown file. The file is timestamped (e.g., consult-2026-03-30.md) and contains the full reasoning, frameworks applied, and recommendations. A concise summary is returned to the conversation with a pointer to the file.
This means you always get both: a quick answer in the chat, and a complete document you can share, reference, or build on.
All agents can be invoked via slash commands or triggered through natural language. Multi-turn agents (most of them) support resume — you can push back, ask follow-ups, or continue where you left off.
The agents chain together, but the workflow adapts to where you are. Start by describing your concept and what materials you have — Claude will assess your stage, propose which agents to run (and which to skip), and create a task list to track progress. Every step is optional. A founder with a pitch deck and interview data gets a different plan than someone with just an idea.
See the using-product-kit skill for full workflow documentation with example plans by stage.
The interview coach and summary agents require a transcript file. Place the transcript in your working folder and reference it by path:
/coach Evaluate the interview at ./transcripts/customer-interview-01.md
Reference textbooks are fetched automatically from GitHub at runtime.
This plugin uses the agents/ directory pattern from the Claude Code plugin spec. Each agent is a markdown file with a system prompt body. In Cowork, agents are launched via the claude CLI with --model sonnet to bypass the forced Haiku subagent model, ensuring complex analysis runs on a capable model.
Deliverable-first: Every agent writes a complete markdown file to ./outputs/. You always have a tangible artifact to review, share, or iterate on. This solves the sub-agent context isolation problem — full analysis is always accessible, not hidden in a separate context window.
DRY routing: All 16 command files are thin stubs (~25 lines) that reference the central SKILL.md for routing logic. CLI flags, model selection, and launch patterns are defined once in SKILL.md and inherited by every command.
Heartbeat protocol: Agents report progress via lightweight heartbeat files (./outputs/.heartbeat-{agent}.json), updated at phase transitions. A shared protocol doc (docs/heartbeat-protocol.md) is injected into every agent via --append-system-prompt-file, so the parent Claude can track progress without polling the result file.
Context externalization: Agents read files via the Read tool and fetch references via WebFetch — they never expect content to be injected into their launch prompt.
Resume support: Multi-turn agents (critic, PRD builder, persona developer, etc.) maintain full context across rounds. Ask follow-ups naturally.
The YC Review and CEO Review agents incorporate frameworks from gstack by Garry Tan. These frameworks are adapted for Cowork (gstack is designed for Claude Code) by extracting the pure conceptual content and wrapping it in our own role, voice, and deliverable sections.
A nightly GitHub Action (sync-gstack.yml) automatically:
- Fetches the latest
office-hoursandplan-ceo-reviewSKILL.md files fromgarrytan/gstack - Extracts clean framework content (strips bash preamble, telemetry, gstack-specific tooling)
- Compares against our stored snapshots
- If changes are detected, merges the new framework into our agent files (preserving our custom sections via
<!-- GSTACK-FRAMEWORK -->markers) and opens a PR
The sync pipeline has a 31-test TDD suite covering extraction, merge, diff detection, and end-to-end simulation.
-
gstack by Garry Tan (MIT) —
/yc-reviewand/ceo-reviewbuild on Garry's office-hours and plan-ceo-review frameworks. -
BMAD Method by BMad (MIT) — The Critic's 6-pass adversarial document review and the five stress-testing techniques in
/vc-review(First Principles Decomposition, Reverse Brainstorming, Six Thinking Hats, Red Team vs Blue Team, Analogous Company Analysis, Debate Club Showdown) are influenced by BMAD's structured elicitation and analysis workflows. -
Jonathan Ellis / Sandalphon Capital — The
/vc-reviewagent's gated screening pipeline (Elevator Clarity, Problem Severity, TAM, Timing Catalyst, Delta 4, Problem Decomposition, Solution-Problem Fit, Competitive Positioning, Pre-Mortem) and BMAD-based adversarial stress-testing architecture are adapted from Jonathan's automated healthcare VC diligence system, generalized for cross-industry use. -
Alexander Osterwalder & Yves Pigneur — Business Model Generation (Wiley, 2010). The
/bizmodelagent's nine-block Business Model Canvas structure, misalignment diagnosis, and business model pattern language. -
Larry Keeley, Ryan Pikkel, Brian Quinn & Helen Walters — Ten Types of Innovation (Wiley, 2013). The
/bizmodelagent's Doblin framework pushes founders past product-only thinking across all ten innovation types: Profit Model, Network, Structure, Process, Product Performance, Product System, Service, Channel, Brand, and Customer Engagement. -
Madhavan Ramanujam & Georg Tacke — Monetizing Innovation (Wiley, 2016). The
/pricingagent's framework: four monetization failures (Feature Shock, Minivation, Hidden Gem, Undead) and the 9 Rules of Monetization. -
Board of Innovation — 50+ Business Model Examples. The
/bizmodelagent's company-anchored pattern library and "What if?" provocations.
See CHANGELOG.md for the full release history.
MIT — see LICENSE.
John Renaldi — jrenaldi79 | presto.consulting