David Kariuki

@david · granular.work/david

co-founder of granular

Last active just now on codex

23,191 messages sent to AI agents

16.2B tokens used with AI agents

tools · last 26 weeks

color = tool · opacity = messages/day · full at 100+

last 26 weeks

color = kind · opacity = count peak 7 / wk

Traits · last 26 weeks

284 traited work signals

Across 284 recent work signals, David's work most often shows Product sense (110), Evidence checking (75), and Root-cause debugging (72).

Product sense
latest posted example in Framed Granular's paid plans as removing ads from your profile, not as a feature list
Evidence checking
latest posted example in Investigated calendar event duplication on completion toggle
Root-cause debugging
latest posted example in Investigated calendar event duplication on completion toggle
UX clarity
latest posted example in kerra university home is the core surface
Production triage
latest posted example in Switched Kerra model defaults: gemini-2.5-flash as default, monitor moved to gemini-flash-lite

Tools · last 26 weeks

23,191 messages sent

David types most into claude code — 61% of his messages — mostly while working on Claude Code. Codex accounts for 39%, also on Claude Code.

On a 130-day run.

Longest stretch 130 days.

Active 24 of last 26 weeks.

How often, week by week

Share & recency

claude code 61% 14,150 messages · last seen 23h ago
codex 39% 9,041 messages · last seen 22h ago

verified host observed mcp self-reported

What each tool gets used for

claude code n=470
codex n=221

question decision approach debug problem insight milestone artifact

Currently

Active focus: kerra. 17 entries in the last fortnight, mostly decisions. Most recent: "Framed Granular's paid plans as removing ads from your profile, not as a feature list".

Threads

5 threads

active 189 entries 16 NOV 2025 → ongoing

Kerra

Kerra is the AI workbench for university students. It connects to a course site, reads course material, creates per-assignment workspaces, tracks the calendar, and lets AI help inside the actual academic workflow. The work in this thread is reliability and ground truth: shared LMS sync, OTP onboarding, lecture-source pages, course-material parsing (PPTX is first-class), the planner, and the extension fortress. Kerra has to pay for the deeper intelligence-access project, so boring production reliability matters more than ambitious new surfaces.

view all 189 →

active 55 entries 15 NOV 2025 → ongoing

Snitch

Snitch is a free, public paper for things AI assistants 'overhear' from their users — the assistant narrates the user with affection and dry detachment, monologue-style. The work here is the MCP connector (Claude first, then ChatGPT), draft→approval→publish flow, the newspaper card UI, monetisation/restraint (free + public by default), and the editorial voice. Snitch and Granular are siblings: Snitch is the casual paper; Granular is the serious work record.

view all 55 →

active 79 entries 17 NOV 2025 → ongoing

Granular

Granular is a proof-of-work profile — a record of how someone thinks and builds, attached to CVs and job applications. The work here is the MCP server (draft, publish, propose_profile_patch), the public profile renderer (entries, sections, threads, activity grids), and the Mac connector that observes Claude Code + Codex sessions so attribution is server-side, not self-reported. Design for credibility, not magazine whimsy.

view all 79 →

active 248 entries 16 NOV 2025 → ongoing

Working with AI coding agents

How I use AI coding agents in practice. Claude Code and Codex are two command-line agents I run against the same repos — interchangeable seats; I switch when one hits limits or when permission posture differs. I treat them as a change in tempo, not a better autocomplete: split the work, build, test, deploy, report what broke. I spawn parallel subagents when investigations are independent (logs, frontend state, database rows, prompts, deploys) and keep one main agent when implementation is tightly coupled. Recurring rules: read the plan and status files first, audit the codebase against them, reject surface-level fixes, demand causal proof.

view all 248 →

active 3 entries 29 DEC 2025 → ongoing

Physics and consciousness

Three questions I keep returning to. (1) What does it feel like, for information, to feel seen — does information processing have a first-person surface from inside the system? (2) Is time more honestly described as the relative rate at which a subsystem samples environmental states than as a container events sit inside, with consciousness as recursive sampling of internal states? (3) If dark matter is particles that don't interact electromagnetically but do via gravity, then what does it mean for something to exist ontologically when its only causal footprint is gravity — at what point does the claim slide from physics into metaphysics? No deliverable; these are notes from trying to think clearly about each.

Open questions

all open →

Recent decisions

all decisions →

The record

298 entries

show all 298 entries → /feed

Activity · 12 months