How rohunvora/just-fucking-cancel works
This repo is not a traditional app with runtime services (no backend, no DB, no deployable API). It’s a Claude Code skill: a packaged set of instructions + assets that teach Claude how to run a local workflow:
• Ingest local bank/credit-card CSVs
• Detect recurring charges
• Interactively classify them (Keep / Investigate / Cancel)
• Generate a local HTML audit report
• Optionally perform browser automation (open Chrome, navigate cancellation flows)
Skill package (the “product”)
• .claude/skills/just-fucking-cancel/SKILL.md
The core “program” in this repo: instructions, prompts, step-by-step logic Claude follows.
• .claude/skills/just-fucking-cancel/assets/template.html
The report template Claude uses to generate the audit HTML.
• .claude/skills/just-fucking-cancel/references/common-services.md
Domain knowledge / reference data Claude uses for classification, heuristics, known merchants, etc.
Local data inputs / examples
• data/subscriptions.json
Likely example data, defaults, or a structure the skill reads/writes (even if used as reference).
• User-provided: “exported .csv’s to your local folder” (from README)
Docs / onboarding
• README.md, GETTING_STARTED.md, getting-started.html
Hosting/deployment (minor/optional
• vercel.json
User (Operator)
• Provides CSV exports
• Answers classification questions
• Confirms cancellations
Claude Code (runtime environment)
• The actual execution context where the skill runs
• Has access to local filesystem and can drive tools (browser)
Just F*ing Cancel Skill
• As a subcomponent inside Claude Code or as a plugin attached to it
• Internals:
• “Skill Instructions” (SKILL.md)
• “Reference Knowledge” (common-services.md)
• “Report Template” (template.html)
Local Filesystem
• Input: CSV statements (user drops in)
• Output: Generated HTML report (local file)
• Optional: Intermediate data (e.g., normalized transactions, subscription list JSON)
Report Viewer (Browser / Local HTML)
• User opens the generated HTML report locally
• Contains checkboxes, copy button, privacy toggle, collapsible sections (from README)
Browser Automation Controller
• Claude-driven automation that opens Chrome and performs navigation/clicks
External Subscription Provider Websites
• Netflix/Spotify/Gym/SaaS cancellation pages
• As a cloud/external boundary
Anthropic Claude Model (Cloud)
• The model that processes conversation/instructions
• Important for the privacy/trust boundary: the README explicitly distinguishes “local CSV stays local” vs “conversation reaches Anthropic”.
Browser Automation Controller ↔ External Websites: “HTTP interactions + UI automation”
10. Browser Automation Controller ↔ User: “Confirm before final steps / user can intervene”
Claude Code ↔ Anthropic Claude Model (Cloud): “Prompts, reasoning, dialogue”
12. A note/annotation: “Raw CSV not uploaded; user data remains local” (this is a core product claim in the README, so it belongs visually).
Architectural patterns/principles to highlight (specific to this repo)
This is best diagrammed as a “local-first, AI-orchestrated workflow”:
• Instruction-driven orchestration: SKILL.md is effectively the “application logic”
• Reference-driven classification: common-services.md and maybe subscriptions.json act as knowledge sources
• Template-based report generation: template.html is a deterministic output structure
• Human-in-the-loop: user Q&A is a first-class control mechanism (not an edge case)
• Trust boundaries:
• Local filesystem + browser actions are on the user’s machine
• Model inference occurs in the cloud (Anthropic)
These deserve explicit visual treatment (boundaries, annotations).
Concrete labeling/coloring conventions (so it’s readable and unambiguous)
• Gray cloud: external services: subscription websites
• Orange cloud: Anthropic Claude Model (explicit trust boundary)
• SKILL.md as “Orchestration Logic”
• assets/template.html as “Presentation Template”
• references/common-services.md as “Merchant/Service Knowledge Base”
• data/subscriptions.json as “Seed Data / Example Catalog (optional)”
• vercel.json as “Static docs hosting config (optional)”
Small callouts next to relevant edges:
• Next to Local FS ↔ Claude Code: “CSV stays local”
• Next to Claude Model cloud: “Only conversation/context sent”
• Next to Browser Automation → Websites: “Human confirmation before final cancellation”
• Next to Report HTML: “Checkboxes + privacy blur + copy cancel list”
These are key product properties described in README and should appear visually.
That set will be the “best and most accurate” representation for this repo because it captures: instruction-based logic, local-first data handling, human-in-the-loop decisions, and optional browser automation across trust boundaries.