How I Use Claude Code to Build Upwork Proposals That Convert

For most of the time I have been on Upwork I sent the same thing every Upwork attorney sends: a cover letter. Three or four paragraphs of credentials and a flat fee at the bottom. It works fine. It also looks identical to every other proposal in the prospective client's inbox.

The proposals I lost most often were not the proposals where my pricing was wrong or my credentials were thin. They were the proposals where the prospect could not tell, from a paragraph of prose, what I was actually going to do for them in the first 24 hours, what the risk picture looked like, what they were going to receive at the end, and how the engagement would unfold past day two. They could not tell because cover-letter format does not give the reader a way to see those things at a glance. So they hired the attorney who told the simplest story, which was often the cheapest one.

The change I made is structural. Instead of describing my engagement in paragraphs, I render it as a one-screen interactive dashboard the client can scan in 90 seconds and read in detail in five minutes. The same dashboard answers the four questions every prospective client actually asks before hiring: what will you do first, what is the exposure, what will it cost, and what happens past the first deliverable. The dashboard is built in HTML attached to the Upwork message, so the client opens it in their browser, scrolls, and clicks through the same way they would scroll through any product page. It looks like a project workspace, because it is one.

I write each dashboard with Claude Code. I do not have it write proposals for me. I have it help me build them faster than I could alone, with a consistent architecture across matter types, and with the supervision protocols a California attorney is required to maintain over any AI assistance.

Each module answers a different question the prospective client has not yet asked out loud. I do not always include every module on every proposal. A short consultative engagement will skip the multi-phase architecture; a large defense engagement will lead with it. But these are the building blocks.

Plus an Engage card at the bottom (Upwork primary CTA + paid-consultation secondary CTA) and a footer with the bar-number link and attorney-advertising disclaimer. Those are not numbered because they appear on every proposal regardless of matter type.

This is the actual sequence I run when a new Upwork invitation lands. The first time I built a dashboard from scratch it took most of an afternoon. With the modules and an established CSS toolkit, the current cycle is about an hour from invitation to send. Most of that hour is me reading and revising, not the AI generating.

This is what one of these dashboards looks like in compact form. The fictional client is Acme Roasters, LLC, a hypothetical small California coffee-roaster facing a demand letter from a former co-packer alleging missed monthly minimums under a 2024 supply agreement. The client, the matter, the documents, and the numbers are entirely invented for this example. I am showing the architecture, not a real engagement.

The full version of this proposal would also include the sticky module navigation, the Practice Fit credentials tiles, the FAQ accordion, the Research Library, and the Engage CTA card. I trimmed those here so the example fits on one screen. Every module on a real dashboard is collapsible or expandable so the client controls how much detail they want at any moment.

When I was iterating on the dashboard format inside Claude Design (the canvas product), the strongest variation it surfaced was an information-architecture pass, not a visual-polish pass. The fix for "wall of text" turned out to be folding the six dense sections behind one-line previews and giving the skim-reader a hero band, a three-cell glance card, and a CTA above the fold. The variant below is a compact version of that idea applied to the same fictional Acme matter. The same folding pattern is also what I just applied to the rest of this page.

Credit: the folded-iteration pattern (hero band + 3-cell glance card + 6 collapsed sections with status tags and one-line previews) is borrowed from a Claude Design canvas pass on this same fictional matter. The information-architecture insight (the fix for "wall of text" is folding, not more visual polish) is the part worth stealing. The same pattern can be ported to any of the dashboard variants.

The California Rules of Professional Conduct and the November 2023 California Bar Practical Guidance on Generative AI apply to my proposal workflow the same way they apply to my brief-writing workflow. ABA Formal Opinion 512 (July 2024) is the same standard nationally. The mechanical practices I follow:

• I do not feed prospective-client confidential information to the model before engagement. The job post, sample documents shared publicly through the Upwork interface, and the matter type are all I work with at proposal stage. Conflict-and-scope clearance happens before any privileged documents enter the workflow.
• I verify every citation. If the dashboard cites Greenley v. Kochava or Henry Schein v. Archer & White, I have either personally read the case or it appears in a hub page I authored after personally reading the case. I do not let the model cite cases I have not read.
• I never let the model promise outcomes. Every framing is "illustrative pressure estimate," "preliminary framework," "depends on documents." If the model drifts into outcome-language during drafting, I rewrite it.
• I run a verb scan on the final draft. Every "I will" sentence has to match scope I am actually committing to. This catches the most common AI failure mode: enthusiastic over-commitment dressed up as professional language.
• I review every word before send. No proposal goes to a prospective client without me reading it from top to bottom. The AI accelerates the production; it does not replace the supervision.
• I disclose AI assistance when asked. If a client asks whether I use AI in my practice, I say yes and walk them through this exact workflow. The dashboard format is itself a transparency artifact: the architecture of how I think about a matter is visible to the client before they hire me.

This page exists so prospective clients and other attorneys can understand what I do. It is not a sales pitch for an AI product, and it is not a template you can pick up and run with by yourself unless you have already built the rest of the supervision protocol around it.

• It is not "AI writes my proposals." The AI drafts the scaffolding. I review every word and rewrite anything I am not personally confident saying under my bar number.
• It is not generic. Each dashboard is built for the specific matter type and the specific prospective client. The fictional Acme Roasters example above looks structurally similar to a real dashboard, but every line on a real one is drafted for that engagement.
• It is not a shortcut around competence (RPC 1.1). If I do not understand the matter type at the level the dashboard implies, the dashboard does not save me. The hub pages I cite are research I did over years before any AI was involved.
• It is not free of error. AI assistance occasionally introduces a citation issue, a date issue, or an over-promise. That is why every draft goes through a human review pass before any prospective client sees it.
• It is not the right fit for every practice. A solo or small firm that handles five matters a year does not need this. The break-even on building a dashboard library is somewhere around twenty to thirty proposals a year, and the format works best on matter types where the prospective client benefits from seeing the architecture before hiring (defense engagements, complex transactional work, multi-phase projects).