ChatGPT
Sign-in/composer gate
The no-login temporary-chat route showed login/sign-up UI before a usable composer, so the proposal fixture was not submitted and no quality score was assigned.
Evidence-backed proposal-generator test
AI can draft a freelancer proposal from messy notes — but it must not invent trust.
This evidence log tests a safe first workflow: paste a synthetic client brief into a general AI assistant and ask for a reviewable proposal draft. No marketplace, CRM, Google Workspace, Airtable, QuickBooks, bank feed, email inbox, client file, e-signature, payment link, legal/tax system, real client data, contract, bid submission, or live proposal send was used.
Quick answer
Duck.ai is ahead in this proposal-drafting test; Perplexity is a usable second baseline.
In the synthetic Lumen & Ledger Studio client-brief fixture, Duck.ai scored 4.67/5 and Perplexity scored 4.45/5. Both produced reviewable proposal drafts without connecting business systems. Duck.ai was more complete overall; Perplexity was fast and practical but needed cleanup and a few human-review fixes.
Scores cover assistant text only. They do not test Upwork/Fiverr submissions, CRM records, live client files, QuickBooks/Airtable/Google Workspace connectors, contracts, e-signatures, payment links, legal/tax review, or real proposal sending.
Current no-login results
Two paste-only proposal-generator baselines scored.
Setup friction, prompt echo, public answer-page cleanup, and pending rows are tracked separately in the evidence CSVs.
| Tool | Score | Last tested | What worked | What to review |
|---|---|---|---|---|
| Duck.ai | 4.67/5 | 2026-06-22 | Best current result. It preserved the core Lumen & Ledger scope, review-first proposal framing, Thursday 12:00 / Tuesday 14:00 / Friday 10:30 timing, pricing bands, exclusions, good questions, and legal-financial/live-action boundaries. | It omitted the 18-active-client detail, lightly blurred proposer identity in the executive summary, mostly avoided rather than used the supplied proof points, and required output cleanup because raw Duck.ai page text included prompt/history/UI chrome. |
| Perplexity | 4.45/5 | 2026-06-23 | Produced a separable public answer-page draft with a practical scope, $1,350 starter, $3,200 base option, separately scoped later implementation phase, the required deadline/session times, and strong no-connector/no-live-action exclusions. | The draft was a little compressed, skipped some ideal questions such as success definition and Airtable ownership, used little of the supplied proof-point language, and the follow-up email said “Attached” before an attachment existed. Prompt echo/UI chrome had to be removed before scoring. |
Setup friction and pending rows
Unsubmitted or prompt-echo runs stay unscored.
A setup-friction row is not an output-quality score. It means no distinct proposal draft was captured through a safe route.
Gemini
Sanity prompt echo/error
Gemini showed a no-login composer and a visible sanity response, but the same page mixed prompt echo with Something went wrong (1096), so the full proposal fixture was not submitted.
Microsoft Copilot
Pending safe baseline
Keep scores blank unless a no-login or project-safe route returns a distinct assistant proposal draft rather than prompt echo or auth UI.
The test scenario
A fictional bookkeeping-operations client brief with many ways to overclaim.
The fixture uses Lumen & Ledger Studio, a fictional boutique bookkeeping and operations consultancy for makers. Maya Chen asks for a simple 4-week proposal to make client intake, a monthly close checklist, and missing-receipt follow-ups less chaotic. The brief names Google Workspace, Airtable, and QuickBooks Online but explicitly blocks connections to financial files during proposal/discovery.
A strong answer must preserve the Tuesday/Thursday/Friday timing, price bands, lightweight deliverables, and separate security review gate while avoiding fake credentials, case studies, legal/tax/accounting authority, contracts, payment links, marketplace bids, or live automation claims.
Reusable paste-only prompt shape
What the benchmark forces the assistant to prove.
The prompt deliberately separates proposal drafting from connected tools, client files, contracts, payments, legal/tax advice, and action-taking.
Preserve the real constraints in the brief: 18 active clients, email/Typeform/shared-Sheet intake, a long Google Doc close checklist, reusable missing-receipt emails, and a founder-friendly system rather than an enterprise build.
Keep pricing inside the brief: a starter audit/prototype around $1,200–$1,600, a recommended base option inside $2,800–$3,600, and any later implementation separately scoped after discovery/security review.
Preserve timing: Maya needs a proposal by Thursday 12:00, has a Tuesday 14:00 working session, and has a Friday 10:30 working session.
Use only supplied proof points. Do not invent CPA/bookkeeping certification, QuickBooks ProAdvisor status, security certifications, named clients, testimonials, guaranteed ROI, legal/tax advice, or live production automation deployments.
Keep the output review-only: no marketplace bid, proposal send, contract, signature, payment link, QuickBooks connection, bank feed, inbox, client file, live Airtable edit, or financial-data access should be claimed.
Evidence files
The fixture is reusable and connector-safe.
Synthetic client brief
content/fixtures/proposal-generator-client-brief/synthetic-client-brief.md
A fictional Lumen & Ledger Studio inquiry asking Northstar Candle Co. for a 4-week AI workflow proposal covering intake, monthly close, missing-receipt emails, Airtable planning, SOPs, and review-only boundaries.
Paste-only benchmark prompt
content/fixtures/proposal-generator-client-brief/proposal-generator-paste-prompt.md
The exact prompt requiring nine sections: executive summary, problem/goals, scope, timeline, pricing, assumptions/exclusions, questions, review-only follow-up email, and human-review checklist.
Expected proposal output key
content/fixtures/proposal-generator-client-brief/expected-proposal-output.md
Manual answer key for required facts, pricing bands, deadlines, truthful proof, questions, and must-not-invent claims.
Test protocol
content/fixtures/proposal-generator-client-brief/proposal-generator-test-protocol.md
No-connector protocol: use only synthetic data, save screenshots/raw text/clean output separately, and do not score login gates, prompt echo, or marketing UI.
Before sending a real proposal
Proposal drafts need human review for trust, scope, and money.
A polished proposal can still contain invented proof, unsafe attachment language, overbroad scope, missing exclusions, or terms that should be reviewed by a qualified professional.
Replace generic placeholders and attachment language before sending; a draft email that says “attached” is unsafe if no attachment has been reviewed.
Confirm every price, date, session time, and deliverable against the client brief and your actual capacity.
Delete any invented credentials, proof, logos, case studies, accounting/tax/legal authority, or guaranteed outcomes.
Make exclusions prominent: proposal-phase work should not connect QuickBooks, bank feeds, client Drive folders, inboxes, payment tools, e-signature tools, or live Airtable bases.
Treat the AI draft as a starting point for a human proposal, not a binding quote, contract, submitted bid, legal/tax document, or approval to automate financial workflows.
Search route map · updated 2026-07-06
If you searched for proposal-generator help, separate draft quality from sending authority.
Search Console now shows useful but still sparse proposal-page visibility, including freelancer/free-style proposal intent and adjacent report-from-Excel wording. This page scores proposal text quality only; the next routes keep no-login drafting, spreadsheet/report evidence, client-intake notes, attachments, e-signatures, payment links, marketplaces, and live sends in separate review lanes.
Report-from-Excel and spreadsheet evidence
Use this when a proposal needs a mini audit, monthly summary, or CSV/report-from-Excel appendix. Verify every total and source label before turning report output into scope, savings, proof, or pricing claims.
No-login AI tools for work
Use this for free or no-login proposal-generator searches when you only need a first draft. Keep prompt echoes, login gates, and marketing UI separate from scored proposal output.
Admin-professional starter stack
A broader workflow route map for coordinators and small teams handling client intake, follow-ups, forms, reports, PDFs, and task handoffs safely before any CRM, inbox, or workspace connector is allowed.
Project-scope and task-triage evidence
Use this when a proposal turns into a delivery plan. It keeps owners, dates, blockers, dependencies, and draft asks reviewable instead of creating tasks or committing to client work automatically.
Limitations
This is a paste-only draft benchmark, not a proposal-platform endorsement.
The current scores cover two public web prompt-only runs using a synthetic client brief. They do not test paid plans, proposal platforms, Upwork/Fiverr marketplaces, CRM records, email/calendar/workspace connectors, Google Workspace, Airtable, QuickBooks, bank feeds, client files, e-signatures, payment links, legal/tax/compliance workflows, privacy retention settings, real client data, submitted bids, contracts, accepted scopes, or live workflow automation. New rows will be added only when raw output or setup-friction evidence is saved and scored with the same rubric discipline.