# How Indragie Karunaratne actually uses coding agents

Context, uniprof, Sentry AI lead

**If you take one thing:** `Claudescope`, transcript link in every PR. he built his own hook-driven tool so reviewers read the agent session and plan behind a PR instead of the generated diff.

- **Agent:** Claude Code (cli) ([source, 2026-04-11](https://github.com/indragiek/GHFS/commit/80afde292d219f4f033d1f683dcaaae016ed877a))
- **Agent:** Codex ([source, 2026-04-11](https://github.com/indragiek/GHFS/blob/fe08755bcc9120d2bf339531f982b98a29225c04/agents.toml#L2))
- **Model:** Claude Opus 5.5 ([source, 2026-10-12](https://x.com/indragie/status/2109671763277975935))
- **Model:** GPT-5 ([source, 2025-09-22](https://www.indragie.com/blog/i-shipped-a-tool-to-help-agents-fix-slow-code))
- **Review:** Reviews the plan and prompts, not the 10,000-line diff ([source, 2026-04-20](https://www.youtube.com/watch?v=IPBIB7Go6TY&t=237s))
- **Harness:** Claude Code in the terminal, GitHub Desktop for diffs ([source, 2025-07-01](https://www.indragie.com/blog/i-shipped-a-macos-app-built-entirely-by-claude-code))
- **Rules file:** Per-project CLAUDE.md = AGENTS.md, driven by a lat.md knowledge graph ([source, 2026-04-11](https://github.com/indragiek/GHFS/blob/fe08755bcc9120d2bf339531f982b98a29225c04/CLAUDE.md#L8-L14))
- **Skills:** dotagents, swiftui-pro, swift-testing-pro, swift-concurrency-pro, lat-md, sosumi ([source, 2026-04-11](https://github.com/indragiek/GHFS/tree/fe08755bcc9120d2bf339531f982b98a29225c04))
- **Hooks:** PostToolUse hook: uploads the session transcript on git commit / gh pr create ([source, 2026-04-20](https://www.youtube.com/watch?v=IPBIB7Go6TY&t=416s))
- **MCP servers:** XcodeBuildMCP ([source, 2026-04-11](https://github.com/indragiek/GHFS/blob/fe08755bcc9120d2bf339531f982b98a29225c04/agents.toml#L11-L14))

## In their words

> Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

> agents = ["claude", "codex"]

> I had Opus 5.5 use captured HTTP traffic from the mobile apps for Ring (cameras), Petlibro (feeder), and Whisker (litter robot) to build an hour-by-hour summary

> I've been writing (generating) more code, trying new tools (hello GPT-5 and Codex), and trying to push the boundaries of what agents can do.

> you don't want to review a 10,000 line slop PR but you may be interested in reviewing like the plan that was used to implement that which is significantly smaller than the code output


Archetype (our label): Toolsmith. Newest source: 2026-10-12.
Card: https://www.prehistoriq.com/stacks/indragiek/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/indragiek.json