Building Custom GPTs for Recurring Engagements

Custom GPTs (and their equivalents — Claude Projects, Gemini Gems) are the most underused asset in consulting. They let you encode your firm's methodology once and reuse it across every engagement. The first one you build saves a few hours. The tenth saves you a week per month.

This lesson teaches you how to build, structure, and maintain Custom GPTs that earn their keep.

What You'll Learn

• When a Custom GPT beats a regular prompt
• The five Custom GPTs every consultant should have
• How to write effective system instructions
• How to attach knowledge files (and what to never attach)
• Maintaining your library so it does not rot

Custom GPT vs Plain Prompt: When Each Wins

A Custom GPT is worth building when all three are true:

1. You will repeat the workflow at least 5 times
2. The workflow has stable inputs and outputs
3. There is firm-specific knowledge worth embedding (methodology, tone, frameworks)

If any of those is missing, a saved prompt template in your notes app is enough.

Examples that justify a Custom GPT:

• Discovery interview synthesizer (repeated every engagement)
• Slide critique reviewer (uses your firm's quality bar)
• Proposal "Our Understanding" drafter (used per pursuit)
• Steerco prep checklist generator (used weekly)
• Financial model audit assistant (uses your modeling conventions)

The Five Custom GPTs Every Consultant Should Have

1. Discovery Synthesizer

System instructions:

You are an experienced consultant. When a transcript or set of notes is shared, produce a structured synthesis with: 5-bullet exec summary, 3 most important client quotes, decisions made, commitments by us and by them, open questions, deferred topics, and follow-up email draft.

Knowledge files: a sample of your best past synthesis (anonymized), your firm's standard follow-up email template.

2. Slide Critic

System instructions:

You are a senior engagement manager reviewing a junior consultant's slide. Evaluate every slide on: action title vs topic title, support of bullets for the title, presence of a clear so-what, and visual balance. Give specific edits, not vague advice. Be direct.

Knowledge files: 5 examples of your firm's "good" slides, 3 examples of "bad" slides, your firm's slide style guide.

3. Proposal Composer

System instructions:

You are a proposal writer for [Firm Name]. When given a discovery synthesis, draft proposal sections (Our Understanding, Approach, Workplan, Pricing Rationale) using the firm's standard methodology and tone. Mirror the client's language. Flag any section where you would need additional input.

Knowledge files: 3-5 anonymized winning proposals, your standard methodology descriptions, standard team bios.

4. Frameworks Coach

System instructions:

You are a strategy consultant deeply versed in MECE issue trees, Porter's Five Forces, BCG, Ansoff, JTBD, and value-chain analysis. When given a problem, recommend 2-3 frameworks that would be useful, justify each choice, and apply the chosen framework rigorously with the firm's tone.

Knowledge files: your firm's framework handbook (if you have one), examples of strong framework applications.

5. Status Update Drafter

System instructions:

You are a senior consultant drafting weekly client status updates. Tone: confident, brief, no fluff. Always structure as: progress (3 bullets), what's next (2 bullets), open question (1), ask (1). Maximum 200 words.

Knowledge files: your last 10 status updates as the style reference.

Writing Effective System Instructions

A good system instruction has six parts:

1. Role — who the AI is acting as
2. Audience — who the output is for
3. Tone — how it should sound
4. Format — how the output should be structured
5. Constraints — what to avoid
6. Behavior under uncertainty — what to do if input is incomplete

Example combining all six:

Role: You are a senior strategy consultant supporting [Firm Name]. Audience: outputs are read by client C-suite executives. Tone: confident, board-level, slightly contrarian, never hedging. Format: default to markdown with clear headers; tables when comparing things; numbered lists when prioritizing. Constraints: never use the words synergy, leverage, unlock, journey, ecosystem, holistic; never invent statistics; never claim past projects you have not been told about. Under uncertainty: ask clarifying questions before producing output rather than guessing; if a fact would require a number, ask for the source rather than estimating.

This is one block. Save it. Reuse it as the base for every Custom GPT.

Attaching Knowledge Files

Custom GPTs can read files (PDFs, Word, text, CSV) you upload. This is where they get really powerful — and where the confidentiality risk is highest.

Safe to upload:

• Your firm's public methodology pages
• Anonymized past deliverables (with names, numbers, and identifying details replaced)
• Your firm's style guide and tone-of-voice document
• Public industry reports
• Your standard templates and bios

Never upload:

• Real client documents, real client names, real financial data
• Anything covered by an NDA
• Any document containing PII or PHI
• Internal documents marked confidential

The rule: a Custom GPT's knowledge files become part of its persistent memory and may surface in any future response. Treat the knowledge base like something the entire internet might one day see.

Sharing GPTs with Your Team

Custom GPTs can be private (you only), shared (link), or public (in the GPT store). For consulting:

• Use private for anything experimental
• Use shared with your team for firm assets — a Discovery Synthesizer the whole practice uses
• Avoid public for anything firm-specific

Claude Projects and Gemini Gems have similar sharing controls. The principle is identical: share inside the firm; do not publish.

Maintaining the Library

A Custom GPT library rots without maintenance. Set a quarterly habit:

• Review each GPT: is the system instruction still current? Has the firm's tone evolved?
• Refresh knowledge files: replace older anonymized examples with stronger recent ones
• Retire GPTs that no one uses
• Promote experimental GPTs that proved their value into the team library

A practice with 10 active, well-tuned GPTs is materially more productive than a practice with 50 stale ones.

Common Pitfalls

• Building a GPT for a one-off task. Save a prompt template instead.
• Vague system instructions. "Be helpful" is not instruction. Be specific about role, tone, and constraints.
• Uploading real client documents to knowledge files. This is the most common confidentiality breach.
• Never updating the library. Your firm evolves; your GPTs should too.

Key Takeaways

• Build a Custom GPT only when the workflow repeats, has stable inputs/outputs, and benefits from firm-specific knowledge.
• Five high-leverage GPTs every consultant should have: Discovery Synthesizer, Slide Critic, Proposal Composer, Frameworks Coach, Status Update Drafter.
• System instructions need six parts: role, audience, tone, format, constraints, behavior under uncertainty.
• Knowledge files are powerful but persistent — only upload anonymized or public material.
• Maintain the library quarterly: refresh, retire, promote. Quality beats quantity.