Proposal Writing 2 - Custom GPT Limitations – Axum

💡

Understand the current limits of Axum’s Proposal GPT—required inputs, missing knowledge packs, industry‑wide caveats, and the human QA steps you must take

📍 Step 1: Confirm Prerequisites (What the GPT Needs to Work)

Before asking Axum Proposal GPT for any draft, make sure you have these 7 core inputs (our “seven tenets” for this GPT):

Client name

Sector

Geography

Opportunity type (RFP/EOI/unsolicited)

Due date (YYYY‑MM‑DD)

Required sections & word counts (as stated by the client)

Team members to include (names; confirm bios/CVs)

Also provide:

RFP/brief (uplod or paste link or text excerpts)

Budget info (roles, # people per role, days per person, and Rate Card A–D)

✅ Tip: If any of the seven inputs are missing, ask the GPT to flag gaps before it drafts.

🚧 Note: No personal data in screenshots. Redact client‑sensitive info in examples.

📍 Step 2: Current Axum Knowledge Gaps (What’s Still Missing)

As of 2025‑10‑06, the Axum Proposal GPT is still being onboarded with internal knowledge. Expect these gaps:

Knowledge Packs: Incomplete. We haven’t loaded all historical proposals, credentials, or boilerplates. Outputs may miss key Axum precedents or reuse generic wording.

Sector Packs: Not yet available (work in progress). Sector‑specific frameworks, indicators, and sample paragraphs are partial or absent; drafts may lack the sector depth you expect.

Relevant Experience Library: Incomplete. The GPT does not know all past Axum projects or results. You may see thin or generic examples until we add more cases.

Team Bios/CVs: Partial coverage. If names aren’t provided (or bios not loaded), profiles may be missing or outdated or generalized and not suited to a particular sector

Template Nuance: It follows the Axum proposal template, but client‑specific formatting (e.g., page limits, numbering styles, annex rules) must be explicitly provided in the prompt.

Evidence Citations: It can cite public facts, but access to internal sources is limited; some citations may need manual verification.

Automations Not Wired: It cannot auto‑fetch files from internal drives, CRMs, or email. You must paste/upload the RFP/brief and any proprietary inputs.

🚧 Note: Where sector packs are missing, instruct it to list assumptions and data to confirm at the top of the draft.

📍 Step 3: Industry‑Wide Limitations of Proposal GPTs

These apply to most custom proposal GPTs—ours included:

Context fragility. If key inputs are vague, models produce generic boilerplate (risks: sameness, weak differentiation).

Hallucination risk. May invent stats, program names, or over‑claim Axum credentials unless forced to cite and verify.

Citation brittleness. Inline brackets can be correct in format but factually off if the source wasn’t properly retrieved or is outdated.

Formatting fidelity. Models struggle with strict RFP formats (exact page counts, annex labels, specific tables). Expect manual layout work.

Numerical accuracy. Budgets and totals can be mis‑summed without a spreadsheet check; FX, VAT, and per diems are not handled automatically.

PDF parsing issues. Scanned or complex RFPs may be misread; sections can be dropped or reordered.

Version control. The model does not manage versions or approvals; treat each chat as context‑scoped.

Data governance. The GPT won’t know confidentiality levels; humans must ensure NDAs and client IP are respected.

No system actions. It can’t submit on portals, obtain signatures, or schedule reviews.

✅ Tip: Ask it to produce an “Evidence Plan” (5–10 facts + sources) before drafting.

🚧 Note: Treat outputs as first drafts—not final client copy.

📍 Step 4: Consulting‑Specific Caveats & Risks

Where consulting proposals are unique, watch for:

Differentiation & POV. Without sector packs, the narrative may lack Axum’s distinct POV (Afrocentric, coalition‑building, institutional capacity). Add this explicitly from context and client work we have that you want to anchor the proposal

Claims about experience. The model can misattribute results or imply engagements we didn’t deliver. Cross‑check every project claim.

Compliance & eligibility. It does not validate eligibility criteria, procurement clauses, or legal requirements (anti‑corruption, sanctions, beneficial ownership).

Budget realism. It doesn’t know internal availability, utilization targets, write‑off policies, or fee caps; partner/finance review is required.

Staff bios. Without current data, bios may omit promotions, new certifications, or language capabilities.

Client tone & politics. It can’t sense org politics or stakeholder sensitivities; BD Lead must tune tone and emphasis.

Consortium roles. It won’t automatically allocate responsibilities across partners or flag conflicts of interest.

Localization. Country policy details (tax, regulatory constraints, local per diems) require human confirmation.

Page limits & annex math. The model estimates pages; it won’t enforce exact counts across narrative + annexes.

✅ Tip: Instruct it to map each RFP requirement to a draft section (a compliance matrix) so humans can verify coverage.

🚧 Note: PM, SPM and Partner‑level sign‑off is mandatory before client circulation.

📍 Step 5: Quality Gate & Required Human Review

Run this human‑in‑the‑loop check every time:

Structure & compliance: Does the draft mirror RFP headings/page limits?

Evidence & citations: Are stats current and sourced? Replace weak/uncited claims.

Experience claims: Confirm client names, dates, locations, outcomes; remove anything uncertain.

Voice & POV: Ensure Afrocentric, systems‑thinking framing with local institutions and coalition design.

Budget math: Recalculate totals, reimbursables, VAT treatment, FX, and rate card selection (A–D).

Risks & assumptions: Add a one‑page “Choices We Made” and to‑do list for the PM.

Sensitive content: Check NDA terms, confidentiality labels, and required consents for named partners.

Who signs off:

BD Lead: Storyline, compliance matrix, executive summary.

Sector Lead: Technical accuracy, evidence, comparators.

Finance/Partner: Rate card, totals, VAT/FX, commercial terms.

PM: Team availability, timeline feasibility.

(Insert Screenshot: Quality Gate — Sign‑off Roles)

✅ Tip: Ask the GPT to generate the compliance matrix and assumptions/risks page—then you finalize.

🚧 Note: Never send a draft to a client without human edits.

📍 Step 7: Risk‑Mitigation Checklist (Copy/Paste)

Use this exact prompt with Axum Proposal GPT, then complete the human checks:

Model prompt

Before drafting, list (a) an Evidence Plan with 5–10 facts + planned sources,
(b) a Compliance Matrix tying RFP requirements to sections, and
(c) a Data Gaps list. Then wait for my confirmation.

Human checks

RFP pasted (or linked) with page/word rules

Seven inputs provided (client, sector, geography, type, due date, sections/word counts, team)

Sector module specified (or set to unified mode)

Budget inputs (roles, headcount, days, Rate Card A–D)

Evidence verified (replace weak or outdated sources)

Experience claims confirmed with the engagement owner

Math audited (fees, reimbursables, VAT/FX) in a spreadsheet

Voice & POV edited for Axum distinctiveness

Confidentiality & NDA terms checked

PM, SPM and Partner‑level sign‑off recorded

✅ Tip: Keep this checklist in your proposal working doc as the first page.

🚧 Note: Save a PDF of the compliance matrix in the proposal archive for auditability.

✨ Quick Summary

✅ The GPT requires the seven inputs + RFP + budget info to perform.

✅ Knowledge packs, sector packs, and experience library are incomplete; expect thin spots.

✅ Industry‑wide model limits: hallucinations, formatting, math, citations, PDF parsing, versioning.

✅ Consulting specifics: verify eligibility, claims, budget realism, localization, consortium roles.

✅ Human Quality Gate and partner‑level sign‑off are not optional.

🔗 What’s Next?

How to Use Axum Proposal GPT to Draft Proposals Fast (4 Core Use Cases)