Felipe F Garcia

I build the tools that keep AI agents honest.

Fifteen years shipping software, now spent on a narrower problem: an agent's plan is accurate the day it is written, and never again. So I make plans that prove themselves.

Install trigpoint Read the book, free

15+

Years shipping

364

Pages, given away

MIT

Open source

AI

Where the work is now

01 / The problem

The problem trigpoint solves

"The plan did not fail because it was wrong. It failed because it stopped being true, and nothing announced that it had."

trigpoint, a Claude Code plugin

Counts are typed by hand

A header that reads "12 of 40" is one person's memory of the checkboxes below it. Correct the day it is written, wrong soon after. The fix is not more discipline. The fix is to stop typing it.

"Done" is a claim, not a fact

A ticked box asserts that something happened. It carries no evidence that it did. An agent in a hurry ticks the box it believes is finished, and belief is not the same as having run the command and read the output.

Discovered work has nowhere to land

The most valuable thing a week of implementation produces is the work nobody predicted. It gets mentioned in conversation, and it dies when the session ends.

02 / What a run produces

Boxes tick on proof, not on assumption.

This is not a mock-up. It is trigpoint's own ROADMAP.md as it stood on 2026-08-27, whose drift gate runs against it in CI.

ROADMAP.md 27 tasks, 25 done
  • 1.2 Write ledger validation: a ticked task with no Verified line is an error, a duplicate task id is an error, an unknown Blocked by reference is a warning Verified: python3 -m unittest tests.test_ledger_validate -v -> Ran 18 tests in 0.001s / OK. 2026-08-26 done
  • 1.3 Write scripts/check_drift.py, the read-only CI gate over the same parse, with exit codes 0, 1 and 2 Verified: python3 -m unittest tests.test_check_drift -v -> Ran 7 tests in 0.002s / OK. 2026-08-26 done
  • 6.8 Verify one non-Claude harness by actually installing it. The Codex, Cursor and Gemini manifests are built to each published shape but none has been installed, so they are a claim rather than a fact until one is not started

Nobody typed 25. It is derived from the tasks themselves, so it cannot disagree with them, and CI fails the build if it ever does.

/plugin marketplace add felipeflorencio/claude-plugins
/plugin install trigpoint@felipeflorencio
View on GitHub

03 / About

Fifteen years building software. The last stretch spent on whether you can trust what an agent tells you.

I am a self-taught software engineer. For fifteen years I built other people's systems: backends, mobile applications, platforms, at companies from a handful of people upwards. Nobody gave me a curriculum, so I learned by shipping things and reading the parts that broke.

What I work on now is narrower. I run my own infrastructure rather than renting a platform, I run models locally as well as through APIs, and I write tooling for agent-assisted development. Not the demo layer. The part underneath, where the question is not whether an agent can produce code but whether you can believe its report of what it did.

trigpoint came out of that. It exists because I kept watching good plans quietly stop being true, and because a ticked box that carries no evidence is just a rumour with better formatting.

If you are doing something in that territory and want another pair of hands or another opinion, write to me.

04 / The book

The Ultimate Spring Boot Bundle. Now free.

I wrote it to sell. It sold badly. It is a genuinely good reference and it was doing nobody any good sitting behind a price, so it does not sit behind one any more. From architecture to execution, with the diagrams and the sample code.

  • 3eBooks
  • 2Diagrams
  • 364Pages
  • 0Price

Get the bundle

The download appears here straight away. One email asks you to confirm your address before anything else is ever sent, and that confirmation carries a one-click unsubscribe link.