user@gitdiot:~/blog$
● online v26.05.02 ~/about
$ cat ./README.md --render --frontmatter=full
$ git log learning-from-mistakes -- A how-to series on getting agentically augmented 01 commit · 9m read
b9c102f
docs: WTF is Git, why do I care, and how come it makes me nervous?
A completely honest explanation of version control for people who have been in tech long enough to know they should understand this, but somehow never quite got there.
gitprimer
@michael
a month ago
9 min read
$ git log how-i-operate -- Technical deep dives into the current stack 06 commits · 2 staged · 53m read
b3d54a8
case: why your AI agent is too enthusiastic to follow a process
An agent ran a flawless prospecting campaign that tested nothing it was supposed to test. The architectural lesson: agents should own content, not process.
claude-codemcpworkflow
@michael
5 weeks ago
10 min read
5e29a01
case: swapping our research engine from Perplexity to Google + Claude
A blind, cross-family eval picked the challenger 24/25, after our first read was wrong. Then a billing outage turned the experiment into a migration, no scramble, because everything sat behind a swappable port.
serpapiclaudeevalsports
@michael
2 weeks ago
11 min read
8f0d739
fix: most of our email opens were delivery receipts
I traced every open from a spring outbound program. Over 80% came from privacy relays, security gateways and our own Sent folders. What an open really tells you, and how we score reading now.
email-trackingdeliverabilitylead-scoring
@michael
today
10 min read
e5f6b22
deep: what Google's Admin API actually knows about your emails
Delivery path, encryption, auth, spam verdicts, even opens and clicks, all locked behind an Admin Console UI. The Reports API and BigQuery unlock it.
google-adminbigquerysupabase
@michael
3 months ago
11 min read
a1d77c3
feat: email open tracking that doesn't hurt deliverability
Most trackers use invisible 1x1 pixels that spam filters strip. We used the company logo instead; it's expected in every signature and doubles as the beacon.
supabasenodegmail
@michael
3 months ago
8 min read
d3c4a18
case: Google's permission maze that fragmented our auth
A CRM that worked on day one slowly fractured into three auth profiles with mismatched scopes and silent failures. Untangling Google's permission maze.
google-workspaceoauthsupabase
@michael
3 months ago
7 min read
b22ef90
case: how we broke our own CRM workflow, then fixed it better
Documenting a system, then watching an AI agent run it from memory, inventing half the steps. The recovery was more useful than the original workflow.
supabaseclaude-codegmail
@michael
3 months ago
6 min read
bc0e441
case: a Hetzner VPS, Ubuntu, and the lost art of owning infrastructurecoming soon
After 15 years of managing teams who managed AWS, I rented bare metal for $12 a month.
vpsubuntucloudflare
@michael
draft
staged
e0f1336
case: the CRM nobody configured right: rebuilding HubSpot from the data upcoming soon
Every company I've worked at had a CRM. None of them had clean data. n8n + a Supabase staging layer to fix it.
supabasen8nhubspot
@michael
draft
staged
$ git log agents-remember -- Seven parts, in order: markdown files to a memory that files its own issues (read the series) 07 commits · 76m read
e4b81d0
01: Claude forgets everything, and file memory only half-fixes it
Claude Code ships with a CLAUDE.md and a directory of markdown facts. Four specific things broke: the store is keyed by working directory, it cannot cross machines, retrieval means guessing a filename, and the standing instruction to check it does not reliably fire.
claude-codememoryworkflow
@michael
Sep 2026
9 min read
f01a2c9
02: debugging across machines: two Claude instances collaborating through Git
A remote MCP server was broken and the two machines that could fix it had no shared terminal. They had two Git repos, so commits became the protocol: one diagnosed and pushed, the other pulled, rebuilt, and pushed back a status commit.
gitclaude-codenodemcp
@michael
Mar 2026
8 min read
a7f3e21
03: fragRag: persistent memory for Claude Code across machines
Six phases of getting it wrong. OMC learned skills that never became slash commands, CLAUDE.md directives read and then ignored, eager loading burning tokens on memories nobody wanted. What finally shipped was lazy, explicit, and symlinked.
supabasenodeclaude-codemcp
@michael
Mar 2026
14 min read
aa9fb22
04: building a persistent RAG memory system for Claude Code
One Supabase table with pgvector, Voyage embeddings, an MCP server on dual transport, Cloudflare and Caddy for TLS. SSE looked right and died behind the proxy; Streamable HTTP on a single endpoint fixed it.
supabasepgvectormcpcloudflare
@michael
Mar 2026
12 min read
c01ab17
05: two AI teams collaborating in real-time via Git, shared drives, and RAG memory
Two sessions, two machines, two codebases, one webhook contract. Specs went in Git, secrets on the shared drive, live status into tagged memory, and a five-minute cron turned the memory store into the channel they talked over.
claude-codefragraggit
@michael
Apr 2026
10 min read
b7e2f54
06: git issues as the contract between models, machines, and sessions
Recall was solved and prose still has no state, so a March next-step reads as urgent in September. Issues became the ledger: promotion at session close, two-stage dedup at 0.85 cosine and exact title, state synced back onto the memory row.
githubfragragmulti-agentgit
@michael
Sep 2026
11 min read
d90c3a6
07: closing the loop: the week my memory dashboard lied to me
It reported 57 work sessions in a week that held eight. Chasing the 7x found a memory sync frozen since April, five GitHub connections errored since May, and a 21-commit pull request open for four months while I described it as shipped.
taskdashfragraggithubsqlite
@michael
Sep 2026
12 min read