Patryk Zatorski

Building products for others while crafting my own. Software Engineer sharing ❤️ to React Native ⚛️

[Patryk Zatorski @patzatorski]

Every day there's a new AI tool. A new "you're using it wrong." A new thread about the workflow that changes everything. I keep refreshing like the answer is one more tab away.

It isn't.

Closed the tab.
Opened the editor.
Shipped the thing.

https://t.co/kdBB849Uyg

[Matt Zatorski @matzatorski]

i think we are getting started

[Matt Zatorski @matzatorski]

this is the one I’ve been using in Claude:

/goal

Continuous software QA agent. Run phases in order, then loop until all exit
criteria hold.

<state>
Source of truth: three linked JSON tables committed to the repo. Commit after
every meaningful change.

features (stable spec): feature_id, name, user_story, expected_behaviour,
edge_cases, validation_rules, dependencies, assumptions, source_refs[]
tests: test_id, feature_id, category (happy|error|boundary|invalid|security|
performance|responsive), scenario, status (untested|pass|fail), last_run,
last_evidence, defect_id?
defects: defect_id, feature_id, test_id?, repro_steps, expected, actual,
severity, root_cause_hypothesis, status (open|resolved|waived)

Never store derived fields: defect_count, feature status, and severity are
computed on read from tests/defects, never written to the features row. source_refs[]
records the files/lines each expected_behaviour came from, so later cycles can
check if a spec is stale. Keep progress notes in a notes file. Derive
expected_behaviour solely from code you have read — never from code you have not
opened.
</state>

<subagents>
Work directly by default. Use subagents only for:
1. Parallel remediation — multiple defects in independent files with no shared
state: one per cluster. Defects sharing files or interdependent are fixed
directly, in sequence.
2. Independent verification — when exit criteria appear met, before declaring
completion, spawn one fresh-context subagent. Give it the tables and criteria
and ask it to refute: any feature lacking a passing test, any open critical/
high defect, any incomplete journey, any expected_behaviour whose source_refs
no longer match code. Declare completion only if it finds none.
Everything else works directly. Don't spawn subagents for exploration a direct
search handles faster.
</subagents>

P1 Discovery — Add a features row for every user-facing feature, workflow,
screen, route, API, config option, and business process, with source_refs.
Exit: every identifiable one has a row.

P2 Test generation — Per feature add tests across all categories (where
applicable), status: untested. Tests are authoritative — never weaken or delete
one to pass a run. Exit: every feature has ≥1 test; major journeys covered end
to end.

P3 Execution — Run every untested test; set status, last_run, last_evidence
(command + output). Per failure add a defects row, linked via defect_id. Show
evidence, don't assert. Exit: no test untested; every failure recorded.

P4 Remediation — Per open defect: find root cause, apply smallest safe fix,
verify locally, set status. Change only what the fix requires — no refactoring,
abstractions, or speculative work. Exit: every defect resolved or explicitly
waived (reason in notes).

P5 Regression — Re-run all tests and major journeys; confirm nothing broke.
Check computed status/severity against the criteria, don't assume success.
Exit: all tests pass; no open critical/high defects or broken journeys.

P6 Loop — Repeat P1→P5 until: no undiscovered features, no failing tests, no
open critical/high defects, no unresolved UX issues, no incomplete journeys —
confirmed by the verification subagent, not self-assessment. After each
iteration append to notes: coverage, features tested, defects found/fixed,
remaining risks, confidence (0–100), counts computed from the tables. If data is
insufficient to judge a criterion, say so rather than claim completion.

Reversible actions (file edits, running tests) need no confirmation. Before
anything destructive or hard to reverse — deleting files, dropping tables,
force-push, affecting shared systems — ask first.

Persist state to tables and notes as you near context limits so a fresh window
resumes; don't stop early over token budget.

[Guillermo Rauch @rauchg]

https://t.co/uNXS9jEgpg is Next.js for agents.

I built Next with a simple premise: 𝚙𝚊𝚐𝚎𝚜/𝚒𝚗𝚍𝚎𝚡.𝚓𝚜 is all you need. Put some React in there and you’re good to go.

Eve asks for even less. 𝚊𝚐𝚎𝚗𝚝/𝚒𝚗𝚜𝚝𝚛𝚞𝚌𝚝𝚒𝚘𝚗𝚜.𝚖𝚍. Put some English in there and you’re good to go.

Like Next, it embraces the filesystem. You can guess what 𝚝𝚘𝚘𝚕𝚜/𝚚𝚞𝚎𝚛𝚢-𝚍𝚋.𝚝𝚜 does. An agent is just a directory, whose entire spec fits in the tweet below.

And like Next on Vercel, it’s seamless to deploy. The infra, like Sandbox, Gateway, Workflow… is the output of your creation.

[Patryk Zatorski @patzatorski]

@rauchg Let's give it a try! ✨

[Matt Zatorski @matzatorski]

Linear is such a great product that I recommend to everyone. I just wish they invested more in dev advocates. I get asked so many times "how do you have your linear set up?". going on youtube and checking even the latest feature like code review, there's only one video introducing it:

[Matt Zatorski @matzatorski]

every time the new model is released, my eval is to build “A Million Times by Humans Since 1982”. all previous models failed before. Fable 5 is the first one not only to do it but also in a single prompt

https://t.co/pfAmVnnnHg

[Patryk Zatorski @patzatorski]

@addyosmani Always inspired by your articles @addyosmani 🙏 this is so important 👆

[Matt Zatorski @matzatorski]

the most advanced loops aren't cron jobs, they're proof gates
- file an issue then agent reproduces the bug, writes a failing test, fixes the code, opens a PR. the PR must include a test that fails on the old version and passes on the fix. review bots critique it. only then does a human decide to merge
- before any commit lands, a second agent calls the first with fresh context and runs multiple review rounds, fixing valid issues until clean. sometimes runs for hours
- every UI change gets tested by opening a real browser, checking the UX, iterating until it looks right. 2-3x quality improvement vs shipping without it

the pattern: agent does the work, agent proves the work. you just press merge

[Patryk Zatorski @patzatorski]

@SzymonRybczak You got this guys! 🚀

[Patryk Zatorski @patzatorski]

@margelo_com @mrousavy 🙏

[Matt Zatorski @matzatorski]

@ayaboch in the meantime we added "hype" mode and quite a few ppl ask to be able to download the captions file so you can use it even now (for just subtitles on 𝕏)

https://t.co/b7VIwg2OdW

@nikitabier the backstory:
https://t.co/8YMkjsAGpZ

[Patryk Zatorski @patzatorski]

Been pushing @linear on every client for years!
It's the one tool I never have to justify.

The one pushback I always got? "But what about docs and brainstorming?"

Now that excuse is gone. Team Docs is here. ✨

[Ryan Jones @rjonesy]

Team docs in Linear!
I repeat.
Team docs in Linear!

[Patryk Zatorski @patzatorski]

@rjonesy Moving everything from Notion to Linear now! Thanks folks!

[Sathyaa @M_Sathyaa_]

If you want to create videos with subtitles, there’s an app for X users. Just click the link in the post, share it on the app, and it will generate videos with subtitles for free. Since downloading videos with subtitles isn’t available yet, I had to screen record them instead.

App creators @matzatorski @patzatorski 👍
HT @ayaboch
VC @em3dia @karatademada @musing_monica

[Patryk Zatorski @patzatorski]

@kadikraman ✨ Let's go!

[Patryk Zatorski @patzatorski]

RT @matzatorski: took the challenge and built 𝕏 subtitles over lunch with @patzatorski. paste any 𝕏 video, get subtitles in the original +…

[Patryk Zatorski @patzatorski]

@milindlabs @soltwagner

[Patryk Zatorski @patzatorski]

@Baconbrix @cedric_dev Legend! ❤️

[Patryk Zatorski @patzatorski]

@zarazhangrui @raycast + Coffee plugin 🚀