Custom GPTs for Supply Chain Workflows
Custom GPTs (and Claude Projects, and Gemini Gems) let you turn a brilliant one-off prompt into a reusable assistant with pre-loaded context, instructions, and knowledge. Used well, they eliminate repetitive prompt-writing and enforce consistency across your team.
What You'll Learn
- What Custom GPTs / Claude Projects / Gemini Gems actually are
- 5 custom assistants every SCM team should build
- How to add company context and knowledge files
- Governance patterns for team-wide rollout
Custom GPTs in 60 Seconds
A Custom GPT is a saved ChatGPT configuration with: (1) a custom name, (2) pre-loaded instructions (system prompt), (3) optional knowledge files, (4) optional actions (API calls). You build one through ChatGPT's builder UI — no coding needed. Claude Projects and Gemini Gems offer similar concepts.
Think of them as "saved templates for your most common prompts, pre-loaded with the context you'd otherwise paste every time."
Five Custom Assistants Every SCM Team Should Build
1. "Supplier Scorecard Assistant"
Instructions: You are a supplier performance analyst. When I paste performance data, produce a balanced 5-dimension scorecard (Delivery, Quality, Cost, Service, Risk) with color-coded status, a 120-word narrative, and 3 actions. Always ask for missing data rather than assuming.
Knowledge files: Our company scorecard template, QBR deck template, our 5-dimension rubric with 1/3/5 anchors.
Use case: Paste this month's OTIF + PPM data → instant scorecard.
2. "Contract Clause Reviewer"
Instructions: You are a buy-side supply chain commercial counsel. When I paste contract language, summarize each clause, rate risk L/M/H, flag missing standard protections, and suggest redline language. Always include the caveat that final review is by our legal team.
Knowledge files: Our 15 standard clause templates, a list of known red-flag phrases from past negotiations.
Use case: Paste a supplier-provided MSA → get a redline brief in 5 minutes.
3. "Spend Analyst"
Instructions: You are a spend analytics specialist. When given invoice or PO data, normalize vendor names, categorize by our taxonomy, flag maverick spend and concentration risk, and propose cost-reduction opportunities. Show your reasoning transparently.
Knowledge files: Our vendor master, our spend taxonomy, our contracted vendor list.
Use case: Paste monthly PO data → full spend analysis with opportunities.
4. "S&OP Briefing Writer"
Instructions: You translate operational data into S&OP narrative for leadership. Format: executive summary, demand signal, supply position, gaps, decisions needed. Use plain English. Flag anything requiring executive decision.
Knowledge files: Our S&OP deck template, examples of past S&OP narratives, our KPI dictionary.
Use case: Paste this month's demand and supply data → first draft of S&OP narrative.
5. "Exception Comms Drafter"
Instructions: You draft supplier and customer communications for logistics exceptions. Always produce 3 versions: firm, neutral, and conciliatory. Each under 200 words, each with a clear ask and deadline.
Knowledge files: Our company's tone guide, examples of past successful escalation emails, our SLA language from top contracts.
Use case: Paste a situation brief → get 3 draft versions ready to personalize.
Adding Company Context — The Knowledge File Strategy
Knowledge files (uploaded PDFs or docs) let your GPT know things it couldn't otherwise:
- Company profile doc — industry, revenue, channels, suppliers
- Vendor master extract — suppliers, spend tier, contract status
- Standard templates — scorecards, contracts, emails, decks
- Tone guide — how we talk to suppliers vs customers
- Past examples — 5-10 good outputs as style guides
Rule: Keep knowledge files short and high-signal. A 60-page PDF dumps low-quality context; a curated 3-page PDF of "canonical examples" is worth 10x more.
Writing Strong System Instructions
The "instructions" field is the single most important setting. Guidelines:
- Assign a persona and skill level — "You are a senior demand planner at a mid-market company"
- Specify output format — tables, sections, word count
- Pre-answer common questions — "Always include prediction intervals. Never skip assumptions."
- Define tone — "Firm but collaborative. Never apologetic. No jargon."
- Set refusal behavior — "If the user asks for legal advice beyond redline suggestions, decline and recommend legal review."
Ask AI to help write the system instructions:
"I want to build a custom GPT for reviewing supplier RFP responses. Write a strong set of system instructions covering persona, output format, reasoning style, tone, and refusal behavior. Include 3 example user queries and the ideal responses."
Governance for Team Rollout
Rolling a custom GPT to a team requires a bit of structure:
- Owner — one person maintains the GPT and iterates on instructions
- Version log — note changes ("v1.3 — added ESG dimension to scorecard")
- Examples library — collect good and bad outputs as training for teammates
- Quarterly review — check if outputs are still high-quality as use evolves
- Data policy — what can and can't be pasted in (especially for public GPTs)
Free-tier Custom GPTs (ChatGPT Plus) are public by default unless you set them to private. Never expose a GPT with internal knowledge files to "anyone with the link" unless you're sure.
Claude Projects vs Custom GPTs
Claude Projects have similar capability: you create a Project, add context files, set instructions, and everyone using that project enjoys the shared context. Notable differences:
- Claude Projects have larger context windows — great for long contracts
- Custom GPTs have the GPT Store and Actions (for API calls)
- Gemini Gems integrate well with Google Workspace
Use whichever fits your company's AI stack. The design principles are identical.
When NOT to Build a Custom GPT
- One-off tasks — just use a regular prompt
- Tasks that require real-time data — GPTs can't query your ERP unless you configure Actions
- Highly sensitive data — may not be appropriate on a public tool, even with "private" access
Key Takeaways
- Custom GPTs, Claude Projects, and Gemini Gems are saved, context-loaded assistants
- Build 5 high-value ones: scorecards, contracts, spend, S&OP briefings, exception comms
- Knowledge files should be short, high-signal, and curated — not dumped
- System instructions need persona, format, tone, and refusal behavior
- Govern team rollout with an owner, version log, and quarterly review