Product Architect
An open-source product development system covering solo-founder Day 0 to IPO, standing in for the departments you haven't hired yet.
Pure-markdown content system with no external effects, no dependencies, no permission requests; transparent data flow; professional-review disclaimers on legal/financial/HR content and explicit no-fabrication rules. Deducted for: unverified publisher identity, no runtime user-confirmation or isolation mechanism, and disclaimer enforcement relying solely on model self-discipline.
Static cap of 10. Routing tables (SKILL.md + SMART-LOADER.md), context budgets, KDR memory, and troubleshooting sections are self-consistent and detailed with explicit error handling and deadlock protocols. Deducted for: no executed reproduction of key paths, token budgets based on assumptions, scoring/agent-limit rules may not reproduce stably, and no test evidence covering skill outputs.
Audience (solo founder to enterprise), negative triggers (non-product-dev exclusions), and environment fit (Claude.ai/Code/API, free-tier fallbacks) are clearly stated. Deducted for: an extremely broad trigger list spanning nearly all business topics creating real false-trigger risk; no Chinese-language support declared; marketing phrasing ('most comprehensive') sits uneasily with boundary claims.
Excellent layering: SKILL.md entry → SMART-LOADER routing → agent files, with semver CHANGELOG, CONTRIBUTING, MIT license, a repo validator (validate_repo.py) and update path. Deducted for: provided object description still says '31 agents/23 frameworks' vs SKILL.md's '80/36' (version-drift residue); START-HERE.md is truncated in supplied sources; the CI workflow content is not directly verifiable within the reviewed files.
Static cap of 7. Claimed capabilities (PRD, research verdicts, financial models) are backed by example formats and quality standards (Depth Rubric L3+, Enterprise Reasoning Protocol) with clear marginal value over a single generalist prompt. Deducted for: no actual output sample verifying direct usability, no quantified comparative-benefit evidence, and depth promises depending on runtime search-tool availability.
Static cap of 5. Auditable primary material is substantial: repo structure, file counts, change history, and validator-script claims are cross-checkable. Deducted for: actual CI results, the '25/25 compliance checks passed' claim, and the existence/quality of all 80 agent files cannot be independently reproduced in this static review; evidence is single-source (self-describing docs).
- All legal, financial, security and HR output is educational only and requires qualified professional review before real-world use (the skill itself states this).
- The trigger list is extremely broad (pricing to facilities management); false triggering on borderline topics is likely - scope requests explicitly.
- Count claims ('80 agents, 36 frameworks') drift against older descriptions (31/23); users should verify actual files.
- Depth verdicts depend on runtime search tools; without network access, market claims are labeled hypotheses - a white-space verdict is not proof of novelty.
- The skill targets Claude Skills environments only; no Chinese-language support is declared, and India-specific compensation/compliance examples are prominent.
What does this skill do, and when should you use it?
Product Architect is an open-source Claude skill that simulates every department of a product company using many specialized agent files plus supporting frameworks: product, engineering, legal/compliance, finance, marketing, people, and more. It enforces a research-first rule: before recommending you build anything, agents run a market existence/novelty investigation and return a cited verdict. The system controls context budget through a smart loader, resolves conflicts via a cross-agent governance hierarchy, and survives chat compaction using Key Decision Records (KDRs). It fits founders and product managers who lack a full team but need department-head-depth deliverables. Caveat: the README (48 agents/33 frameworks), SKILL.md (80 agents/36 frameworks), and repo description (31 agents/23 frameworks) disagree on scale — verify actual contents after cloning.
On each turn it first reads SMART-LOADER.md to classify the request and route to relevant agent files (max 5 per turn; 3 on free tier); if the request involves building a feature or making a market bet, it runs the Deep Research Protocol first, producing an "already exists (competitors + citations)" or "white-space" verdict; it then loads matching agents (e.g. 04-prd.md with prd-framework.md) to produce PRDs, financial models, security audits, pricing plans, OKRs and other deliverables; every output passes a depth rubric (L3+), a three-layer edge-case doctrine, and a five-level governance hierarchy (Compliance > Privacy > Security > Finance > Chief Reviewer), and each phase emits a structured KDR. Legal, financial, security and HR content is educational and requires professional review.
- A solo founder on day zero validating an idea: ask "does this feature already exist / is it novel?" and get a cited competitive investigation before committing.
- An engineering founder without a PM needs a PRD: trigger "Write a PRD for [feature]" and get happy path, error states, edge cases and acceptance criteria.
- An early team making pricing decisions: load the pricing-packaging framework and pricing agent for price-metric selection, Van Westendorp analysis and discount governance.
- Preparing fundraising or investor updates: use the investor-relations agent and Founder's Playbook for week-by-week costs, fundraising, IP and legal guidance.
- Handling privacy/compliance questions (DSAR, DPO duties): load privacy and compliance agents under the governance hierarchy, with the stricter control winning on conflict.
- Pre-morteming a plan spanning multiple teams or quarters: run the Pre-Mortem Sweep and name 3-5 organisational risks, each with a trigger, owner, 48-hour move and reversal condition.
What are this skill's strengths and limitations?
- Extremely broad coverage: from PRD, MVP and pricing to compliance, fundraising and IPO governance in a single skill.
- Research-first with anti-hallucination rules: agents must never fabricate a company, statistic, study, patent or URL, and must label market claims as hypotheses when no live search is available.
- Engineering-grade design: agent routing for context budget, a five-level conflict hierarchy, and KDR memory that survives chat compaction.
- MIT licensed with no external dependencies; the README claims it works on Claude.ai, Claude Code and API, and offers paste-based workarounds for free-tier users.
- Includes country compliance deep-dives (India, US, EU, UK, Southeast Asia) and 24 SOPs with process maps.
- Self-contradictory scale numbers: the repo description says 31 agents/23 frameworks, the README 48/33, SKILL.md 80/36 — verify actual contents yourself.
- Self-described as an educational framework built via human-AI collaboration; legal, financial, security and HR content must be reviewed by qualified professionals before real-world use.
- The Deep Research Protocol itself admits absence of evidence is not proof of novelty; without live search tools its market verdicts degrade into hypotheses.
- No automated test suite is shown in the repo; validate_repo.py only checks structure, not output quality.
- Context-constrained: 3-5 agents per turn maximum means complex requests require multi-turn phased execution.
How do you install this skill?
Three official routes: 1) Claude.ai (Pro/Max/Team/Enterprise — Skills require a paid plan): on GitHub click Code → Download ZIP, then in Claude.ai → Settings → Capabilities → Skills upload the ZIP. 2) Claude Code: git clone https://github.com/ankitjha67/product-architect.git and place it in your ~/.claude/skills/ directory. 3) API: use the /v1/skills endpoint with the container.skills parameter (requires Code Execution Tool beta). Free-tier workarounds: copy START-HERE.md (Raw) into a normal Claude chat, or paste SKILL.md + SMART-LOADER.md as Project knowledge.
How do you use this skill?
Trigger it with natural language after install, e.g. "Help me build a product", "Write a PRD for a payment feature", "Does this feature already exist?", or "Build a food delivery app for Bangalore" (which runs the phased plan). The system classifies each request and loads only relevant agents (max 5 per turn). Complex requests proceed phase by phase per SMART-LOADER.md; after context compaction, paste the MASTER KDR into a new conversation to restore state. Beyond the documented examples, no further trigger conventions are documented.
How does this skill compare with similar options?
The README positions it as "the most comprehensive open-source product development skill" but names no specific competitor; its Consulting Frameworks borrow from McKinsey 7S, Porter, Blue Ocean and other classics, but there is no directly comparable Claude Skill data in the source.