Next Steps¶
You did it
A claim goes in as three PDFs and comes out settled, or in front of exactly the right person. You designed it from a business document, built it in six proven blocks, hunted ten planted problems, and handed it over pinned, documented and runnable by a stranger.
What you learned, beyond this solution¶
The claims case was the vehicle. These transfer to whatever you build next:
- A complete brief makes a quiet agent. Every answer the planner never had to ask for was an answer you supplied up front.
- One document, one question. Requirements, design, plan: each written once, referenced everywhere, brought to as-built at the end.
- Contracts before consumers. The cheapest half of an interface is the shape: settle it first, and stand-ins can prove the wiring before the real thing exists.
- Gates are commands, and every gate is a floor. Scripts catch the boring half; go and look catches the rest; the agent's "done" is an opinion.
- Restraint is the hard half of automation. A process that flags something on every claim has automated nothing; the clean path settling untouched was your hardest test.
- Write as you go. Context windows fill;
PROGRESS.mdoutlived every session that wrote it. - The screen is the product. Whatever the architecture says, the two reviewer screens are what anyone will ever see; taste is a requirement you state.
- Done means the next person can start: a pinned version, an as-built design, a runbook.
On speed, honestly¶
You just built in a day what the PDD budgeted at 25 person-days, and that first-hand experience is worth more than any statistic. The wider evidence points the same way, with a hedge: as UiPath's CEO puts it, coding agents "compress the first 10% of the work. What's left is the hard part," because "writing code is the easy part. Running it, durably, safely, at enterprise scale, is where the value accrues." Most developers are already there: the Stack Overflow 2025 survey has 84% using or planning to use AI tools. The operating model you practiced today is the part that makes the speed usable: AI proposes. Humans decide. Automation executes. Anthropic's AI-Native SDLC playbook closes on the same note from the other side: "The loop keeps running. Human judgement stays above it." The claims case you built has that shape too: Agents propose, a person decides at the two gates, and the case executes.
Keep building¶
- Keep your toolchain current. Skills and tools evolve quickly:
The CLI and its tools also update themselves on first use each day within the line you pinned, and uip update refreshes the skills. Check uip --version when a build that worked yesterday behaves differently.
- The skills registry is public: fork it and encode your organization's best practices the way this workshop's seed encoded these.
- Start your next build the way this one started: a process document, a generated design, a gate. Do it on your own process, at your own company.
Learn more¶
| Resource | What it covers |
|---|---|
| Coding agents overview | The product docs: start here |
| Working effectively | Official best practices |
| UiPath skills on GitHub | The registry this workshop ran on |
References¶
The works this course quotes, to read in full:
- Daniel Dines, The Work That Remains: Human Judgment, AI, and the Architecture of the Next Enterprise (2026) — download the PDF · UiPath page. The operating model behind the two human gates you built.
- Louis Claxton, The AI-Native SDLC playbook, Anthropic (August 2026) — claude.com. The stage-by-stage view this workshop follows, and the before/after figure in Plan.
- Daniel Dines, From AI speed to enterprise reliability: introducing UiPath for Coding Agents, UiPath blog (May 2026) — uipath.com. Quoted in On speed, honestly.
- Stack Overflow, 2025 Developer Survey: AI — survey.stackoverflow.co. The 84% in On speed, honestly.
Feedback
Found something broken, confusing, or worth teaching differently? Tell your instructor. The findings your agent logged during the build already feed the next edition.