How to use Claude to automate a SaaS coding workflow
A SaaS owner used Claude with a shared Git plan, custom ESLint feedback, Bash approval rules and deterministic workflow generation. They reported lower token spending and shorter workflows while continuing code and UI review.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Turn [feature idea] into structured intent, analyze its impact on [repository] and update [shared plan file] with independently verifiable tasks. Follow [lint and command rules], execute the approved plan, record a Playwright demo of [user flow] and provide the diff and verification evidence for my review.
Reach people exploring AI agents
See placements →What happened
The author reports token spend at about 20% of previous workflows and average workflow time falling from 5–6 hours to 20–30 minutes, while still reviewing most code.
Reported by gojkoa on Hacker NewsHow the agent did it
- Captured structured intent from brief text in a shared Git plan file.
- Mapped intent to the codebase and split the work into independently verifiable tasks.
- Executed work with lint feedback and Bash approval rules.
Where a person stepped in
The author reviews the evolving plan and most completed code, watches recorded UI demos and adds manual exploratory testing when needed.
A person approved each step before the agent continued.
The author runs their own production SaaS and describes a custom Claude workflow built around a single plan file stored in Git. Commands turn brief input into structured intent, map its impact onto the codebase, create independently verifiable tasks, orchestrate execution and support review or demonstration. The author reviews the plan as it develops.
Agent rules are enforced through deterministic tools. Custom linting and style checks run through ESLint after edits, giving immediate feedback. When an agent repeats an unwanted behavior, the author has it write another rule with unit tests. A separate pre-tool regulator parses Bash commands and approves, rejects or requests permission according to its rules.
An orchestrator converts plans into Claude Dynamic Workflow descriptions. The author reports token spending at about twenty percent of previous workflows and average workflow time falling from five or six hours to twenty or thirty minutes. These are their reported measurements. A demo command records UI behavior through Playwright, while static demo pages and API tests provide additional feedback. The author still reviews most completed code through Git diffs and performs exploratory testing when needed. Who it's for: SaaS maintainers willing to invest in repeatable tooling around an agent and retain product and code review responsibilities.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · gojkoa
Frequently asked questions
How were agent rules enforced?
Custom rules and style checks ran through ESLint after edits. A pre-tool Bash regulator parsed commands and applied approval rules, with rejection messages explaining alternatives.
What remained under human review?
The author reviewed the shared plan as it evolved and most completed code through Git diffs. Recorded demos and manual exploratory testing also supported review.
More details
Remaining steps
- Converted plans into workflow descriptions with a deterministic orchestrator.
- Recorded UI demos and reviewed completed code with Git diffs.
About this account
Claude executed planned tasks under deterministic editing and command rules; the author reviewed plans, code and demonstrations.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=49586883. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: gojkoa on Hacker News
Source published 6 Sept 2026
Source checked 14 Oct 2026
Tools and topics
Claude · ESLint · Git · Playwright · regulator script · orchestrator script
SaaS · linting · workflow automation
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.