Custom GPT: An Engagement Methodology and Reporting Assistant

What This Builds

Every new engagement starts the same way: a blank interview guide, a blank scope-of-work section, a report structure rebuilt from memory instead of pulled from what worked on the last three projects. This build turns a library of your past, anonymized proposals, methodology write-ups, and report templates into a Custom GPT that already knows your firm's voice, your standard model-assumption conventions, and your report structure. Ask it for a first draft and get something that reads like it came from your last three engagements, not from a stranger.

Prerequisites

• Comfortable using ChatGPT for drafting and analysis tasks (Level 3)
• Plus subscription ($20/month) to create and use Custom GPTs, at chatgpt.com
• A folder of past engagement documents you're allowed to reuse: proposals, scope-of-work sections, interview guides, report templates, model-assumption memos
• Time to scrub client-identifying details from each document before upload
• Total ongoing cost: $20/month per month (Plus), the only subscription this build needs

The Concept

A Custom GPT works like a new hire who spent their first week reading every proposal your firm has ever written before showing up to their first assignment. You do not re-explain your report structure or your standard assumptions about discount rates and planning horizons every time you start a draft. Configure the GPT once with instructions and a library of reference documents, and every conversation after that starts from what your firm already knows how to do.

Build It Step by Step

Part 1: Prepare your document library

Before touching ChatGPT, gather and scrub five to ten of your strongest past documents:

• Two or three full proposals or scope-of-work sections, with client names, dollar figures, and any identifying details replaced with placeholders like [CLIENT] and [FEE]
• One or two discovery interview guides
• One or two final report templates or table-of-contents structures
• A short memo listing your firm's standard model-assumption conventions: how you typically set discount rates, planning horizons, demand growth assumptions, and constraint framing, written in plain language

Save these scrubbed versions in a dedicated folder. This becomes the knowledge base the GPT draws from, so anything left in these files is something the GPT could surface in a client-facing draft. Do not skip the scrub step for documents you think are "probably fine."

Part 2: Create the Custom GPT

1. Open ChatGPT and look for the GPT builder, typically reached from a "GPTs" or "Explore GPTs" entry point, then a "Create" option. Menu names shift between updates, so look for anything labeled around building or creating a custom assistant
2. Name it something specific to its job, such as "[Firm Name] Engagement Assistant"
3. In the instructions field, paste a system prompt along these lines:

You are a drafting assistant for [Firm Name], a management consulting
practice that does operations research and quantitative analysis work
for [typical client type, e.g., retail supply chain, government agency,
manufacturing] clients.

Your job is to draft engagement documents in the firm's established voice:
- Scope-of-work sections
- Discovery interview guides
- Report structures and section headings
- First drafts of model-assumption sections (discount rates, planning
  horizons, standard constraint framing) based on the conventions in
  your knowledge base

Standard conventions to follow unless told otherwise:
- Planning horizon: [your firm's default, e.g., 3 to 5 years]
- Discount rate: [your firm's default assumption or range]
- Report structure: Executive Summary, Problem Statement, Methodology,
  Findings, Recommendations, Appendix
- Tone: direct, quantitative, no unsupported claims

When drafting, reference the structure and language patterns in your
uploaded documents rather than generic consulting phrasing. If a
request falls outside what your uploaded documents cover, say so
plainly instead of inventing firm conventions that are not documented.

Never include real client names, dollar figures, or identifying details
in any draft. Use placeholders like [CLIENT] and [FEE] unless the user
supplies real values directly in the conversation.

1. Save the instructions

Part 3: Upload your knowledge base

1. In the same configuration screen, find the option to add knowledge files
2. Upload the scrubbed documents from Part 1
3. Save and publish the GPT as private, so only you (or your firm, on a Team plan) can reach it

Part 4: Test it against a real draft

Start a new conversation with the GPT and ask for something you would normally write from scratch:

Check the output against two questions: does it match your firm's actual structure, and does it avoid inventing conventions that were not in your uploaded documents. If it drifts into generic consulting language, that is a sign the knowledge base needs a stronger or more specific example document.

Before any draft goes anywhere near a client, confirm the model-assumption numbers it cites, discount rate, planning horizon, and similar defaults, against your firm's current standard. A GPT built from last year's memo can restate an assumption your firm has since revised, and that is exactly the kind of quiet error a reader would otherwise take on faith.

Real Example: Scoping a Warehouse Consolidation Engagement

Setup: The GPT was built with three past scope-of-work sections, a standard report table of contents, and a model-assumption memo listing a four-year planning horizon and a standard discount rate range.

Input: "Draft a scope-of-work section for a warehouse consolidation engagement. The client currently runs six regional warehouses and wants to know whether consolidating to three or four improves total logistics cost."

Output: A scope-of-work draft with sections for problem definition, data requirements (inbound and outbound volume by warehouse, current lease and labor costs, transportation lane data), modeling approach (a facility-location optimization using the firm's standard four-year planning horizon), and deliverables, phrased in language that matches the uploaded proposals rather than a generic template.

Time saved: A scope section that used to take an hour or more to draft from memory came back as a workable first draft in a few minutes, leaving the remaining time for tailoring it to the client's actual numbers.

What to Do When It Breaks

• The GPT ignores your uploaded documents and drafts something generic → Check file size and count. Very large or numerous files can dilute what the GPT actually references, so trim to your strongest five to eight examples and re-upload
• Drafts start including invented client details or dollar figures → Add an explicit line to the instructions telling it to use placeholders and to ask for real figures rather than inventing them, then re-test
• The voice drifts over a long conversation → Start a new chat for each new document instead of continuing one long thread. Custom GPTs perform best when each request stays close to its uploaded knowledge

Variations

• Simpler version: Skip the Custom GPT builder and keep a single saved prompt with your firm's conventions pasted in, used at the start of any ChatGPT conversation. Less setup, but you retype the reference material every session
• Extended version: Build a second GPT scoped narrowly to model-assumption drafting, separate from proposal writing, so each GPT's knowledge base stays focused and the outputs stay more consistent

What to Do Next

• This week: Build the GPT with five to eight scrubbed documents and run it against one real scope-of-work draft
• This month: Add a model-assumption memo and a report-structure template, then use the GPT for the next full proposal
• Advanced: Pair this with the prompt chain that turns a scoped engagement into a full report, so the same firm conventions carry through from proposal to final deliverable

Advanced guide for Operations Research Analyst professionals. These techniques use more sophisticated AI features that may require paid subscriptions.