claude-procurement-skill

A Claude Code skill that turns any buying decision — from a USB cable to an enterprise SaaS contract — into a professional procurement workflow: formalized requirements, fair-price discovery, total cost of ownership, vendor risk, a documented decision, and a post-purchase lifecycle plan.

---

What it does

• Researches any product — bicycle, PC build, camera, headphones, smart-home device, appliance, cosmetics, furniture, e-mobility, SaaS tool, medical device, or anything else.
• Runs a procurement-grade pipeline modeled on the CIPS 13-stage procurement cycle, condensed to 11 LLM-driven steps.
• Classifies every purchase using the Kraljic matrix (routine / leverage / bottleneck / strategic) so depth scales to importance — no over-engineering a USB cable, no under-engineering a CPAP machine.
• Validates compliance and compatibility before pricing: certification marks (CE, FCC, UL, RoHS, etc.), authorized-dealer verification, return-policy review, IoT cyber/privacy checks, and physical/electrical compatibility with the user's existing ecosystem.
• Discovers fair prices, not just MSRP — eight price tiers tracked (MSRP, street, net, refurb, used, gray market, bundle, trade-in), 3-quote rule, 6–12 month price history via camelcamelcamel / Keepa, peer-tier benchmarks, regional aggregators (Google Shopping, PriceSpy, Idealo, Kakaku, PCPartPicker, and equivalents per region).
• Computes landed-cost TCO over a horizon — consumables, energy, maintenance, expected repairs, insurance net of credit-card coverage, financing, rebates, resale, opportunity cost.
• Surfaces vendor risk — FMEA-lite, single-source-of-failure, ecosystem lock-in, EOL signals, recall history.
• Chooses a decision method that matches the purchase: quick weighted scoring for routine, full weighted matrix for leverage, AHP or Pugh matrix for strategic purchases.
• Stress-tests the decision with a Klein-style pre-mortem and a threshold-based cooling-off pause.
• Plans the lifecycle post-purchase: receive + inspect, warranty registration deadline, day-1 firmware/secure setup, recall monitoring, maintenance schedule, spare-parts stocking, end-of-life resale / recycle.
• Outputs a structured decision memo with full audit trail of sources at tasks/research/<slug>.md. Optionally also emits presentation-ready formats — executive deck (.pptx), polished PDF brief, stakeholder one-pager, or Word .docx — by chaining to the Anthropic anthropic-skills:pptx / pdf / docx skills.
• Token-efficient by design — progressive disclosure + criteria-side caching (self-feeding) + no-re-ask auto-extraction. Saves ~70–80% of tokens and web calls vs naive research, especially across repeat sessions in the same category. See docs/TOKEN_EFFICIENCY.md.
• Calibrated confidence, not vibes — auto-generated domain packs are adversarially verified on high-stakes claims and scored per-dimension (9 dimensions, low/medium/high). The report tells you exactly which data is solid and which to double-check. See docs/QUALITY_MODEL.md.

Why use this instead of "asking Claude"

Asking Claude directly	This skill
One-shot answer	Multi-step, structured pipeline
MSRP anchor	8-tier price discovery + fair-price band (aspirational / target / walk-away)
Sticker price	Landed-cost TCO over horizon, net of CC bonuses + rebates + resale
Generic recommendation	Weighted scoring or AHP / Pugh decision matrix tied to your requirements
No compatibility check	Standards matrix vs. your existing gear
No risk framework	FMEA-lite, vendor risk, single-source-of-failure
Confident-sounding but unchecked	Adversarial verification — high-stakes compatibility / compliance claims must survive an independent refutation pass before they're trusted (details)
"Trust me" confidence	Per-dimension confidence scoring — 9 dimensions each scored low/medium/high, aggregate computed not guessed; the report names exactly which data is weak
No audit trail	Cited sources for every claim, written to disk
No follow-through	Lifecycle plan with dated warranty / recall / maintenance / EOL checklist
One-size-fits-all	Kraljic-driven depth: routine gets a shortcut, strategic gets the full memo
Anglo-centric	Region-neutral; works for US, EU, UK, CA, AU, JP, IN, BR, MENA, and any country
Personal use only	B2C and B2B modes — RFI/RFP/RFQ flow for company procurement
Static knowledge	Self-feeding: derives + saves a new domain pack on first encounter with a category
High token cost per query	Token-efficient by design: progressive disclosure + self-feeding cache (criteria only, never prices) + no-re-ask auto-extraction → ~70–80% fewer tokens and web calls vs naive use across repeat sessions (details)
Markdown report only	Multi-format output: Markdown (default) + executive deck (.pptx) + PDF brief + one-pager + Word .docx, all chained via anthropic-skills:pptx/pdf/docx

Quickstart

1. Clone or download into your Claude Code skills directory:

   cd ~/.claude/skills/   # or wherever Claude Code looks for skills on your system
   git clone https://github.com/<your-handle>/claude-procurement-skill.git

2. Open Claude Code in any working directory (the skill writes outputs to tasks/research/<slug>.md relative to the current dir).
3. Trigger the skill with a natural-language prompt such as:
   • "Research a gravel bike upgrade — I have a Giant Revolt with BSA BB, 142×12 thru-axle, flat-mount brakes, 11sp 105."
   • "Build me a 1440p gaming PC under $1500."
   • "Compare HubSpot vs Pipedrive vs Close for a 25-seat B2B sales team."
   • "Is a SRAM XPLR 12sp rear derailleur compatible with my Shimano 11sp shifter?"
   • "Should I buy a CPAP machine — newly diagnosed with sleep apnea, budget $2000."
   • "Research espresso machines under $1500 for home use."
4. Answer the (single batched) follow-up questions about region, budget, audience level, etc. The skill only asks for what it couldn't auto-extract from your working-directory files.
5. Read the report at tasks/research/<slug>.md. The Executive summary at the top is the headline; the Procurement detail below is the full decision memo.

See examples/ for three fully worked outputs.

How it works — the 11-step procurement workflow

The skill walks every research request through the same pipeline, with step depth scaling to the Kraljic classification from step 1.5:

Pre-purchase

1. Scope & domain — identify product category and intent (new-purchase / upgrade / comparison / compat-check). See references/domain-index.md.
2. Kraljic classification — routine / leverage / bottleneck / strategic. See references/kraljic-classification.md.
3. Load playbooks — procurement-playbook.md, universal-dimensions.md, domain-specific pack.
4. Auto-extract context from working-directory files (gear lists, prior research, notes) — never re-ask for known data.
5. Option-space analysis — make-vs-buy, build-vs-buy-vs-rent-vs-subscribe. See references/option-space.md.
6. Formalize requirements — must-have / nice-to-have (weighted) / dealbreaker. See references/requirements-framework.md.

Sourcing & validation

7. RFI → shortlist + compliance checks — certifications, authorized-dealer verification, return policy, IoT privacy, counterfeit risk, compatibility matrix. See references/compliance-checks.md and references/compatibility-playbook.md.
8. RFP → RFQ + price discovery — MSRP capture, regional aggregator scan, price history, used / refurb floor, peer-tier benchmark, cross-border landed-cost math, red-flag pass, fair-price band. See references/price-discovery.md.
9. TCO + risk + negotiation — landed-cost TCO, FMEA-lite, vendor risk, negotiation tactic playbook. See references/tco-and-risk.md and references/sourcing-and-negotiation.md.

Decision & commit

10. Decide & document — weighted scoring / AHP / Pugh, chosen by Kraljic class. See references/decision-frameworks.md.
11. Pre-commit checks — pre-mortem (Klein), devil's advocate, threshold-based cooling-off pause. See references/pre-commit-checks.md.

Post-purchase (lifecycle)

The report includes a dated checklist for: receive + inspect, warranty registration deadline, day-1 firmware / secure setup, insurance enrollment, recall monitoring, maintenance schedule, spare-parts stocking, end-of-life resale / recycle. See references/lifecycle-management.md.

The exact report shape lives in references/report-template.md.

Domain packs

Pre-built domain knowledge for common categories. Each pack contains the research dimensions, compatibility/standards table, common pitfalls, and trusted sources for that domain.

Domain	Pack	Mode bias
Bicycle	domain-bicycle.md	B2C
PC build	domain-pc.md	B2C / prosumer
Cosmetics (skincare, makeup)	domain-cosmetics.md	B2C
Home appliance (major / white goods)	domain-home-appliance.md	B2C
Small appliance (coffee, blender, vacuum)	domain-small-appliance.md	B2C
Camera (mirrorless, lenses, gear)	domain-camera.md	B2C / prosumer
Audio (headphones, speakers, amps)	domain-audio.md	B2C / prosumer
Smart home (IoT, Matter, Zigbee)	domain-smart-home.md	B2C
Furniture (incl. mattress)	domain-furniture.md	B2C + B2B
Personal mobility (e-bike, e-scooter, EUC)	domain-mobility.md	B2C
IT software (SaaS, business software)	domain-it-software.md	B2B
Medical device (CPAP, hearing aid, CGM)	domain-medical-device.md	B2C + B2B (clinic)
Anything else	domain-unknown.md — self-feeding fallback	—

Self-feeding knowledge base

When you research a category that doesn't have a pack yet, the skill derives the missing dimensions via web search and saves a new domain pack to disk for next time. See references/self-feeding.md.

Auto-generated packs start at confidence: low with an explicit warning banner. After 3+ runs without contradiction, the skill bumps them to medium. When you review a pack and approve it, it gets promoted to human-reviewed / high.

The skill's domain library grows with use, no manual authoring required.

B2C vs B2B modes

The same pipeline serves both personal buyers and corporate procurement. See references/b2b-modifiers.md for the full overlay. Key differences:

• B2C: cooling-off pause based on income threshold; consumer-protection law for returns; personal warranty + insurance considerations.
• B2B: formal RFI → RFP → RFQ process; multi-stakeholder requirements (IT + Finance + Security + Legal + end-user); SOC 2 / ISO 27001 / DPA / BAA validation; AHP for strategic vendor selection; implementation + change-management costs in TCO; approval-chain checklist replacing cooling-off.

Token efficiency

One of the strongest reasons to use this skill over asking Claude directly: structural token efficiency that compounds across sessions.

• Progressive disclosure — SKILL.md is ~180 lines; the 30+ reference modules load only when relevant. A routine-class run pulls 3 files into context; a strategic-class run pulls 6.
• Self-feeding (criteria-side caching) — first encounter with an unknown category runs a one-time deep search across 8 dimensions; result saved as a domain pack with confidence: medium. Two acquisition paths: (a) delegate to the deep-research skill if registered in your Claude Code setup (preferred — fan-out + adversarial verify), or (b) inline 8-category search with parallel-batched WebSearch/WebFetch calls (~20–30 calls) as fallback. Subsequent runs on the same category cost zero web calls for criteria research — only live price discovery + availability is re-fetched.
• Auto-extraction — the skill scans the working directory for existing context (gear lists, prior quotes, prior research) before asking the user anything. Repeated research in the same domain compounds.
• Compressed report template — tables instead of paragraphs; casual mode condenses sections; the canonical Markdown is dense but scannable.

What we cache vs what stays live (this distinction is intentional):

Cached forever (~0 web cost / future run)	Re-fetched every query (always live)
Research dimensions + standards + compatibility axes	Current prices
Certification marks + regulatory bodies per region	Stock / availability
Brand landscape + authorized-distribution patterns	Active promos / discounts
Common pitfalls + failure modes	Fair-price band (calculated per query)
Repairability + EOL norms	Recent recalls / vendor news / FX rates

Caching prices would be a feature anti-pattern — stale prices mislead buyers. The cache is for how to evaluate this category (stable). Live re-fetch is for what it costs today (dynamic).

Rough numbers (illustrative, varies by category):

Scenario	Input tokens	Web calls
Naive "Claude, research X"	40–60k	20–60
This skill, first encounter unknown domain	25–35k	20–30 (deep search) + 5–10 (live prices)
This skill, subsequent encounter same domain	8–15k	3–8 (prices only)

Across 10 runs on the same category, the skill saves ~70–80% of tokens and web calls vs naive use. Full details + verification recipe in docs/TOKEN_EFFICIENCY.md.

Quality & verification

The skill's trust model is explicit and machine-checkable, built on three interlocking mechanisms (full writeup):

1. Structured schemas (references/schemas.md) — every domain pack, report, and in-run object has a defined shape that the skill self-validates before writing. Anti-drift: the 50th auto-generated pack has the same shape as the 1st.

2. Adversarial verification (references/adversarial-verify.md) — high-stakes claims (compatibility, certification, recall status, B2B compliance) must survive an independent refutation pass before they're trusted. The verifier actively searches for evidence the claim is false, uses independent sources, applies four lenses (correctness / conditionality / recency / region), and defaults to unverified, never verified. Prefers delegating to the deep-research skill; falls back to an inline skeptic pass.

3. Per-dimension confidence (references/schemas.md §3) — instead of one blanket label, each of 9 dimensions (standards, compatibility, compliance, brand landscape, pitfalls, repairability, TCO drivers, reviews, buying guides) is scored low/medium/high. The aggregate is computed by a weighted-floor rule, not guessed: a pack can't be "high" while any safety/interop dimension is weak. The report then names exactly which data is solid and which to double-check.

What this buys you: when the skill says a compatibility verdict is verified, an independent pass tried to refute it and failed. When it says a dimension is low, it's flagging genuine uncertainty — not hiding it behind a confident tone. Calibration over vibes.

Example outputs

Three fully worked sample reports demonstrate the skill across modes and Kraljic classes:

• examples/bicycle-gravel-upgrade.md — B2C, leverage class, upgrade intent. Demonstrates compatibility matrix flagging Boost wheels + 12-speed cassettes against an existing 11-speed Shimano 105 + 142×12 thru-axle frame.
• examples/pc-1440p-gaming-build.md — B2C, leverage class, new-build intent. Internal compatibility across CPU ↔ motherboard ↔ RAM ↔ PSU ↔ case; energy line in TCO; weighted scoring across three builds.
• examples/saas-crm-comparison.md — B2B, strategic class, vendor-comparison intent. RFI → RFP → RFQ, SOC 2 / DPA compliance, AHP decision method, multi-year contract negotiation, approval-chain checklist.

Customization

Add a new domain pack

1. Copy the template from any existing references/domain-<name>.md.
2. Fill in: research dimensions, required user inputs, standards & compatibility axes, common pitfalls, trusted sources.
3. Add a row to references/domain-index.md with keywords + status: hand + mode bias.
4. (Optional) Submit a PR — see Contributing below.

Promote an auto-generated pack

When the skill encounters a category for the first time, it writes references/domain-<slug>.md with status: auto-generated, confidence: low and a warning banner.

To promote:

1. Open the file, review accuracy, fill in gaps.
2. Remove the warning banner.
3. Update the frontmatter: status: human-reviewed, confidence: high, human-reviewed: true.
4. Update the corresponding row in references/domain-index.md.

Tune procurement behavior

• Requirements weighting: edit your requirements doc at runtime via the AskUserQuestion prompts; weights live in the report itself.
• TCO horizon: pass 3 years / 5 years / other when the skill asks.
• Cooling-off threshold: ask the skill to use a custom rule (default: 24h pause above 10% of monthly income; 72h pause for strategic-class purchases regardless of spend).
• Source citation strictness: defaults to ≥2 sources for standards/compat claims, ≥1 for opinions, ≥1 aggregator + ≥1 primary seller for price data. Edit SKILL.md "Source citation discipline" section to adjust.

Repository structure

claude-procurement-skill/
├── README.md                          # This file
├── LICENSE                            # MIT
├── SKILL.md                           # Skill entry point + 11-step workflow + routing
├── references/
│   ├── procurement-playbook.md        # Professional-buyer mindset + CIPS mapping
│   ├── kraljic-classification.md      # Routine / leverage / bottleneck / strategic
│   ├── option-space.md                # Make-vs-buy + build/buy/rent/subscribe
│   ├── universal-dimensions.md        # Domain-agnostic evaluation criteria
│   ├── requirements-framework.md      # Must-have / nice-to-have / dealbreaker
│   ├── compliance-checks.md           # Cert marks, authorized dealer, IoT, counterfeit
│   ├── compatibility-playbook.md      # Standards matrix methodology + adversarial verify
│   ├── price-discovery.md             # 8 price tiers, fair-price band, RFI/RFP/RFQ
│   ├── tco-and-risk.md                # Landed-cost TCO + FMEA-lite + vendor risk
│   ├── sourcing-and-negotiation.md    # Negotiation tactic playbook
│   ├── decision-frameworks.md         # Weighted / AHP / Pugh
│   ├── pre-commit-checks.md           # Pre-mortem + devil's advocate + cooling-off
│   ├── lifecycle-management.md        # Receive / register / monitor / maintain / EOL
│   ├── information-sources.md         # Cross-region source catalog
│   ├── self-feeding.md                # How auto-generated packs work + confidence model
│   ├── schemas.md                     # Canonical shapes + per-dimension confidence model
│   ├── adversarial-verify.md          # Refutation methodology for high-stakes claims
│   ├── output-formats.md              # Markdown / .pptx / PDF / one-pager / .docx chains
│   ├── b2b-modifiers.md               # B2C / B2B overlay
│   ├── domain-index.md                # Registry of all domain packs
│   ├── ask-user-patterns.md           # AskUserQuestion question banks
│   ├── report-template.md            # Markdown skeleton for outputs + one-pager template
│   ├── domain-unknown.md              # Deep-search fallback + self-feeding workflow
│   └── domain-<13 packs>.md           # Bicycle, PC, cosmetics, appliances, ... medical-device
├── docs/
│   ├── TOKEN_EFFICIENCY.md            # Why the skill is token-efficient + verification recipe
│   └── QUALITY_MODEL.md               # Schemas + adversarial verify + per-dimension confidence
└── examples/                          # Three fully worked sample reports
    ├── bicycle-gravel-upgrade.md
    ├── pc-1440p-gaming-build.md
    └── saas-crm-comparison.md

Contributing

PRs welcome, especially for new domain packs.

• New domain pack: follow the template structure of an existing pack (e.g., domain-bicycle.md). Include citations for any standards / compatibility claims (≥2 independent sources). Add an entry to domain-index.md.
• Improvements to a process module: keep SKILL.md short — push detail into the relevant references/ file. Maintain region-neutrality (no single-country bias; multi-region examples preferred).
• Bug reports: open an issue with the user query, expected behavior, actual behavior, and which references/ file is implicated.
• Auto-generated pack review: open a PR that promotes an existing auto-generated pack to human-reviewed, with any corrections or additions surfaced from real-world use.

Star history

License

MIT — see LICENSE.

Acknowledgments

The methodology draws on established procurement and decision-science literature:

• CIPS (Chartered Institute of Procurement & Supply) — 13-stage procurement and supply cycle (cips.org).
• Peter Kraljic — Purchasing Must Become Supply Management, Harvard Business Review, 1983 — the supply-risk × importance matrix.
• Thomas L. Saaty — The Analytic Hierarchy Process, 1980 — pairwise-comparison decision method.
• Stuart Pugh — Total Design, 1991 — concept selection matrix.
• Gary Klein — Performing a Project Premortem, Harvard Business Review, 2007.
• FTC 16 CFR Part 233 — Guides Against Deceptive Pricing.
• EU regulations — General Product Safety Regulation, Energy Labelling, EPREL database, Cyber Resilience Act, Right-to-Repair Index.
• Independent testing labs — Wirecutter, RTINGS, Consumer Reports, GamersNexus, iFixit, MPB, KEH, and many more, cited in references/information-sources.md.

Built with Claude Code and the Claude Agent SDK — a skill that researches procurement, built by following procurement-grade research practice.