A Claude Code skill that interviews any business owner about their day-to-day tools, maps their back of house into a visual data architecture, and generates a plain-English 30-day build plan for layering AI on top.
The whole thing rests on one idea: 80% of building an Agentic OS is data prep. The 20% AI layer is easy. This skill is the 80%.
You run /silver-platter inside any Claude Code project. The skill:
-
Silently audits your project folder for existing Claude Code setup (
.claude/,silver_platters/,data/, hooks, agents, rules). If you've already started building, it skips questions about what's there. -
Interviews you in plain English about your tools and workflow. No jargon. It asks what you use, how big your audience/customer base is, and what weekly task you'd most want to automate. It never asks you about "format", "cadence", "volume", or anything that sounds like a database column.
-
Builds a data map using the Pantry / Prep / Plate metaphor:
- Pantry = every place your raw data already lives (Shopify, your inbox, your EHR, your POS)
- Prep Table = weekly summary files ("silver platters") that Claude will write for you
- Plate = the briefs and drafts that land in front of you Monday morning
-
Generates recipes showing how your tools combine for specific business outcomes. Each recipe has: a headline you can screenshot, a "today, manually" movie scene of how you do it by hand, a "next Monday with this live" description, and a time-saved estimate.
-
Generates a 30-day build plan with 5 prioritized steps (conversion hook, silver platters, orchestrator + specialists, audit log + approval gates, slash commands). Each step has "you'll know it's working when" so you can verify yourself.
-
Maps your interaction layer showing WHERE you'll actually read these briefs (Slack, iPad, email, Telegram) and suggests channels you might be missing.
-
Renders a self-contained HTML file (
data_map.html) with 4 interactive tabs you can open in any browser. No external dependencies. Email it, share it, screenshot it. -
Generates OPPORTUNITIES.md listing everything to build, grouped by Claude Code primitive (skills, subagents, hooks, rules, CLIs), with a plain-English glossary so every term is translated.
-
Generates a handoff prompt you can copy-paste to
@claude-code-guideif you want the actual.claude/folder scaffolded for you.
The data_map.html file has 4 tabs:
Cards for every tool in your Pantry, every silver platter in your Prep Table, every brief in your Plate. Click any card for plain-English details including how to connect it to Claude (MCP, CLI, or API with pros/cons for each).
A Sankey-style diagram showing how data moves left to right: raw sources feed silver platters, platters feed briefs, briefs land in the channels you actually use (Slack, iPad, email). If a Pantry box has no line leaving it, that data is sitting unused.
Recipe cards with outcome-first headlines and coral time-saved badges. Click any recipe for the full walkthrough: who triggers it, what Claude does step by step, what the before/after looks like. Each recipe also has a "today, manually" section describing how you do this by hand right now.
Step-by-step setup with: what to do, why this step first, install commands, before/after comparison, and a "you'll know it's working when" sanity check. Step 0 (install Claude Code) is auto-rendered. For regulated archetypes (healthcare, law, wealth advisory), a Bedrock/BAA step is inserted at position 1.
cp -r silver-platter ~/.claude/skills/Then in any Claude Code project:
/silver-platter
- Claude Code installed
- Python 3.9+ with
jinja2(pip install jinja2) - That's it
Runs scripts/audit_existing_folder.py against your project. Detects .claude/CLAUDE.md, settings.json, skills, agents, rules, data/, silver_platters/, outputs/audit_log.md. If anything exists, the interview skips questions about it.
Walkthrough (~30 min, explains every choice) or Fast track (~10 min, assumes you know your tools).
Describe your business in 1-2 sentences. The skill matches you to one of 9 archetypes:
| Archetype | Examples |
|---|---|
ecommerce |
Shopify store, DTC brand, live-stream commerce |
saas |
B2B/B2C subscription software |
professional_services |
Law firm, accounting practice, consultancy |
healthcare_clinic |
Medical practice, dental, specialty clinic |
wealth_advisory |
RIA, hedge fund, family office |
content_creator |
Newsletter, YouTube, podcast, course operator |
restaurant_multilocation |
Multi-unit food service |
real_estate_brokerage |
Agent team, indie brokerage |
local_trades |
HVAC, plumbing, landscaping, cleaning |
If none fit, it falls back to a free-text interview.
The skill asks outcome-shaped questions from references/question_library.md, customized per archetype. It never asks you about file formats, export cadences, or payload sizes. Those are derived silently from references/tool_defaults.md.
Are you using Claude Code already? Any Zapier flows, scheduled jobs?
For high-volume sources, the skill fires plain-English tips: "Six months of Shopify orders is tens of thousands of rows. We'll plan a weekly summary so the agent reads 26 rows instead of 60,000."
Builds the JSON skeleton with auto-derived fields for every pantry/prep/plate item (display names, volume-friendly labels, sample content previews, governing rules).
Walks you through 4-7 starter recipes from references/recipe_templates.md for your archetype. Confirms each one with you. Adds your actual tool names and artifact names. You can add your own.
Builds the 5-step build plan from references/setup_priority_template.md. Customizes each step with your tools, silver platter names, and bot names.
Asks where you want to read your briefs (Slack, iPad, email, terminal). Suggests 1-2 channels you might be missing (e.g., Telegram for time-sensitive alerts).
Runs scripts/render_data_map.py to produce the self-contained data_map.html.
Groups everything to build by Claude Code primitive, with a glossary footer translating every term (skill = a reusable mini-tool, subagent = a specialist staffer, hook = an auto-trigger, rule = guardrails).
Renders a copy-paste prompt for @claude-code-guide to scaffold the .claude/ folder.
Asks explicitly: are you comfortable in Terminal, or would you rather hand this off? If you don't have a developer, the skill generates hire_a_builder.md with an Upwork job posting template, budget ranges ($300-3,000), and the 5 skills a freelancer needs.
Shows you the 4 output files. Offers to teach any piece in plain English or hand off to @claude-code-guide.
After a successful run, your project has:
silver_platter_output/
├── data_map.json # canonical data map (full schema)
├── data_map.html # self-contained 4-tab visualization
├── OPPORTUNITIES.md # plain-English priority list with glossary
└── claude_code_guide_handoff.txt # copy-paste prompt for @claude-code-guide
Every data map this skill produces follows a 4-layer architecture:
Layer 4 — Automation (hooks)
SessionStart: converts messy files on session start
PostToolUse: logs every action to audit_log.md
Stop: blocks exit if a draft is unsigned
Layer 3 — Workers (subagents)
Chief-of-staff orchestrator routes to specialists
Each specialist scoped to one silver platter
Layer 2 — Knowledge (skills + MCPs)
Project-level skills specific to your business
Global skills for tools you'll always have
MCP servers for real-time tool access
Layer 1 — Identity (CLAUDE.md + path-scoped rules)
The foundation. Who your agents are.
What they're allowed to read.
Nested per-folder for domain isolation.
80% of the work is Layers 1-2 (data prep). 20% is Layers 3-4 (the AI on top). Most people build Layer 4 first and wonder why nothing works.
silver-platter/
├── SKILL.md # the skill entry point Claude Code reads
├── references/
│ ├── archetypes.md # 9 business archetypes with pain patterns
│ ├── question_library.md # per-archetype interview questions
│ ├── tool_defaults.md # per-tool schema defaults (format, cadence, volume)
│ ├── recipe_templates.md # 6-10 starter recipes per archetype with full schemas
│ ├── setup_priority_template.md # universal 5-step build plan + regulated exceptions
│ ├── cli_inventory.md # CLI/skill catalog per service domain
│ ├── opportunity_patterns.md # data shape → recommended Claude Code primitive
│ ├── data_engineering_tips.md # plain-English tips triggered by volume/format
│ ├── claude_code_handoff_template.md # Jinja2 template for the handoff prompt
│ └── greenfield_vs_audit_detection.md # how Stage 0 detection works
├── scripts/
│ ├── render_data_map.py # data_map.json → data_map.html via Jinja2
│ ├── audit_existing_folder.py # read-only audit of cwd for existing setup
│ └── templates/
│ └── data_map.html.j2 # the generic HTML template (Pantry/Prep/Plate/Sankey)
└── examples/ # pre-rendered data maps for 5 operators
├── marco_ecommerce/ # Shopify + Meta + TikTok live-stream e-commerce
├── sally_law/ # 12-attorney boutique law firm, Bedrock + matter walling
├── devon_saas/ # B2B SaaS, 280 customers, Codex worktree pattern
├── dr_mehra_clinic/ # healthcare clinic, PHI scoping + controlled bridges
└── dr_anwar_derma/ # solo dermatology practice, Bedrock + biopsy triage
Each archetype has its own question chain, recipe starter library, orchestrator/specialist table, and silver-platter defaults.
| Archetype | Starter recipes | Key Claude Code pattern taught |
|---|---|---|
| E-commerce | 7 (P&L brief, stream prep, refund triage, customer voice, abandoned cart, spend pacing, inventory restock) | Silver-platter pattern, summary tables |
| SaaS | 6 (Monday triage, Codex worktree, weekly product, churn investigation, Slack-to-Linear, release changelog) | Orchestrator + specialists, worktree |
| Professional services | 6 (new matter, time entry, MSJ focus, post-mortem, engagement letter, conflict check) | Path-scoped rules, matter walling |
| Healthcare clinic | 5 (morning brief, lab triage, appeal draft, weekly P&L, referral letter) | Bedrock + PHI scoping, controlled bridges |
| Wealth advisory | 8 (quarterly letter, trade committee, risk check, personal-trade pre-clear, Reg BI doc, client review, billing recon, marketing review) | Bedrock + personal-trade scoping |
| Content creator | 10 (newsletter idea engine, YouTube-to-newsletter, subscriber retention, Gumroad launch, community substitute, creator finance, and more) | Back-catalog as corpus, no-community path |
| Restaurant multi-location | 9 (daily close, food cost, staff schedule, labor variance, aggregator recon, review response, weekly P&L, food waste, manager standup) | Per-location scoping |
| Real estate brokerage | 8 (listing packet, buyer match, closing checklist, lead triage, lead-source ROI, agent scoreboard, dead-deal postmortem, followup rescue) | Per-agent scoping |
| Local trades | 8 (daily dispatch, quote followup, seasonal prep, review response, voicemail triage, weekly revenue, parts restock, maintenance plan) | No-developer path, honest iMessage handling |
The skill explicitly handles operators who have never opened Terminal:
- Questions are phrased in plain English with jargon translated inline on first use
- "I don't have one" is always a valid answer to any "where does X live?" question
- The 30-day build plan includes setup time estimates and "you'll know it's working when" checks
- Stage 9.5 asks whether you're comfortable in Terminal or need to hand off
- If you don't have a developer, the skill generates a hire-a-builder template with Upwork posting, budget guidance ($300-3,000), and the skills to look for
Healthcare clinics, law firms, and wealth advisors get additional handling:
- Bedrock step inserted at position 1 in the build plan (before any data touches the model)
- PHI scoping rule for healthcare (path-scoped, always-on for
data/clinical/**) - Matter walling for law firms (per-matter CLAUDE.md, no cross-matter reads)
- Personal-trade scoping for wealth advisors (client data + pre-clear attestations isolated)
- Controlled cross-domain bridge for the one agent allowed to read both clinical and billing (e.g., insurance appeal drafting), logged on every read
Restaurants with multiple locations, brokerages with multiple agents, clinics with multiple providers:
- The skill detects team size from interview answers
- Per-tenant scoping pattern is surfaced in Stage 5
- Each tenant gets a path-scoped rule (
data/locations/{id}/**,data/agents/{id}/**,data/providers/{id}/**) - The orchestrator sees everything, specialists see only their slice
scripts/templates/data_map.html.j2 is a single-file Jinja2 template that produces a self-contained HTML page with:
- Inline CSS (no external stylesheets)
- Inline JavaScript (no frameworks, no dependencies)
- Three-lane Pantry/Prep/Plate card layout with click-to-modal details
- SVG Sankey diagram rendered dynamically from the data map JSON
- Recipe cards with time-saved badges and modal walkthroughs
- Setup steps with requires/working-when blocks
- Interaction-layer 4th column on the Sankey when
interaction_layeris present - Status badges (green check = have it, gray ? = missing, coral ! = opportunity)
- Sticky legend bar and tab navigation
- Hero stats strip (tools mapped, recipes ready, quick wins, hours/week back)
- Jargon glossary in modals (skill = mini-tool, subagent = specialist staffer, hook = auto-trigger, rule = guardrails)
- Connection method tabs renamed for operators: "Easiest setup" (MCP), "If you live in the terminal" (CLI), "Custom build" (API)
The template renders any persona dict matching the schema. The examples in examples/ are test fixtures demonstrating it.
The full data_map.json schema is documented in SKILL.md Stage 6. Key fields:
business.name, archetype, stack_summary, headline, lead, hours_back_per_week
pantry[].id, tool, format, cadence, volume, feeds, status, cli_skill,
explanation, connection_methods[], volume_friendly, paying_unused
prep[].id, name, domain, sources[], schedule, status, display_name,
domain_friendly, sample_content, governing_rule_excerpt
plate[].id, name, agent, consumers[], approval_gate, reads_from[], status,
agent_friendly, approval_friendly, sample_output, ideation_loop[]
recipes[].id, name, headline, time_saved_per_week, manual_today,
monday_difference, goal, ingredients[], ingredients_friendly[],
claude_code_stack{skills, subagents, hooks, rules},
walkthrough[{actor, action}], before_claude_code, after_claude_code
setup_priority[].step, title, title_friendly, requires, what_to_do, why,
install, before, after, working_when, setup_time
interaction_layer[].id, channel, type, status, description, consumes[]
opportunities[].surface, type, title, explanation, claude_code_feature,
feature_name, estimated_impact
This skill is open-source and free. If you run it against your own business and find a gap (missing archetype, missing tool in defaults, question that confuses non-technical operators), open an issue or PR.
The most valuable contributions are:
- New archetype question chains in
references/question_library.md - New tool entries in
references/tool_defaults.md - New recipe starters in
references/recipe_templates.md - Bug reports from real operators who got confused during the interview
MIT. Use it, fork it, sell services on top of it. Attribution appreciated but not required.