The Problem
Startup founders and product teams routinely pay $5K–$10K for strategy consulting that produces market research, competitive battle cards, and positioning documents. Most of that work is structured, repeatable research and synthesis—not genuine insight. This repo packages that workflow as installable AI agent skills so Claude Code (or any skill-compatible agent) can run the process on demand, for free, in a single session.
What This Does
The repo contains four independent skills, each a folder with a SKILL.md entry point and a references/ directory of detailed prompts. startup-design/ runs a full strategy engagement: market research, customer discovery, competitive analysis, financial projections, and validation experiments. startup-competitors/ produces battle cards, pricing landscapes, and feature matrices from web research. startup-positioning/ applies April Dunford's positioning framework. startup-pitch/ generates investor narratives in multiple formats with scoring and Q&A prep.
Each skill follows the same internal architecture: a multi-wave research loop (typically 3–4 waves), a synthesis step, an honesty protocol that forces negative findings, and a verification agent that cross-checks output. The references/ files are prompt templates that structure each wave's research task.
How It Is Wired
Execution starts when a user invokes a skill—either through natural language (Claude triggers it automatically) or explicitly via /startup:startup-design. The agent reads the skill's SKILL.md, which defines the process workflow and points to references/ files that contain the actual research prompts, output specifications, and quality gates.
There is no code to execute. These are Markdown prompt files that direct an LLM agent to perform web research, synthesize findings, and write structured deliverables. The "call graph" is: user prompt → SKILL.md → references/research-wave-N-*.md → agent performs web searches and writes output files. The references/verification-agent.md in each skill defines a second-pass review step.
Each skill is self-contained—startup-design/ does not import from startup-competitors/, though the README suggests they can be chained manually. The .claude-plugin/marketplace.json file registers all four skills with Claude Code's plugin system.
How To Use It
Setup (from README, verbatim):
claude plugin marketplace add ferdinandobons/startup-skill
claude plugin install startup@startup-skillFor the web app, download .skill files from the Releases page and upload via Settings → Skills.
Configuration: No environment variables or API keys. Skills run entirely through the agent's existing web access and tooling.
Running it: Describe your startup idea or competitive question in natural language. The README documents example triggers: "I want to build a SaaS for real estate agents..." triggers startup-design; "Who are my competitors in the project management space..." triggers startup-competitors. Direct invocation: /startup:startup-design.
Real-World Use
A founder with an idea for a niche SaaS product asks Claude: "Quick validation—fast track mode." The agent runs startup-design in compressed mode, executes the pre-flight check, runs reduced research waves, and returns a go/no-go recommendation with evidence. If the answer is "no," the honesty protocol ensures the agent says so plainly rather than softening the finding.
Code Health & Issues
Static analysis found: No CI/CD pipeline (Medium/SDLC) — no .github/workflows/ or CI config exists; the .github/ directory contains only issue and PR templates. No automated tests run against the prompt files, so regressions in skill behavior are undetectable. The repo has a MIT license and no committed secrets.
The absence of CI is not a functional problem—these are prompt files, not executable code—but it means changes to SKILL.md or references/ files ship without any automated validation that they still produce working skill definitions.
The Bottom Line
This is a well-structured prompt library that productizes a genuine consulting workflow. The honesty protocol and verification agent are thoughtful additions that address the real failure mode of AI-generated strategy: confident but wrong output. For a technical team, the trade-off is that "code health" is really "prompt quality"—you cannot unit-test it, and you must manually verify that prompt changes still produce good results. Suitable for founders who want structured strategy output without the consultant bill, and for agent developers who want a reference architecture for multi-phase research skills.