How to use Codex to build RepoFinder
Christo Roberts used Codex to ship RepoFinder on Cloudflare Workers, verifying contracts in a temporary clone and adding recommendations with visible fallbacks.
letsgochristo.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Build [project-specific recommendation tool] around [data source]. Check current documentation, define verifiable user contracts and implement one shared engine. Reject unsupported candidate identities, label fallback results and run deterministic checks plus judgment evaluations before applying the final changes.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- Christo Roberts
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- RepoFinder live with source and checked-in lessons.
- Human involvement
- Author chose product constraints; separate Claude QA/design tooling found follow-up fixes.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Checked current provider/Cloudflare docs and planned verifiable contracts.
- Built/tested in temporary clone when sandbox blocked repository writes.
- Applied passing diff to original.
Christo Roberts built RepoFinder with Codex to recommend open-source tools that fit a particular project and goal. The deployed application accepts a repository or website, gathers GitHub candidates and explains a shortlist. A browser interface and an MCP tool share the recommendation engine.
Roberts describes a concrete build loop. Codex checked current OpenAI and Cloudflare documentation, planned around verifiable user-visible contracts and encountered a sandbox that could read the repository but not write to it. It made a temporary clone, built and tested there, then applied the diff back after type checking and tests passed.
The implementation separates source extraction from ranking and chat, uses strict structured outputs and excludes repository names that GitHub did not actually supply. If model calls fail, the interface visibly labels a deterministic GitHub fallback. Those boundaries matter because a plausible recommendation is not evidence that a tool exists or suits the project.
RepoFinder was deployed as a Cloudflare Worker with D1 and public source. Separate Claude tooling later supplied QA and design findings. Roberts also notes imperfect evaluation calibration.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Blog · Christo Roberts
Frequently asked questions
How did Codex handle the write restriction?
It built and tested in a temporary clone, then applied the resulting diff to the original repository after type checking and tests passed.
How are invented repository names handled?
Names returned by the model are dropped unless they came from the actual GitHub candidate searches.
What happens when model calls fail?
The app uses a deterministic GitHub-only ranker and displays a GitHub fallback label instead of presenting the result as model reasoning.
More details
Remaining steps
- Deployed TypeScript Worker, D1, browser UI and MCP surface.
About this account
Codex researched current documentation, planned verifiable contracts, built and tested in a temporary clone and applied the checked diff.
Primary first-person account opened and checked on 2026-10-14. Recommendation judgment needs eval calibration; visible deterministic fallback used on API failure. Outcomes and test claims are attributed to the source and were not independently reproduced. Prompt is suggested by Agents In Use, not the author's quoted prompt.
Credit: Christo Roberts on Blog
Source published
Added to Agents In Use
Source checked
Tools and topics
Codex · OpenAI Responses API · TypeScript · Cloudflare Workers · D1 · MCP
Tool discovery · Verification · Cloudflare
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.