Lucas Martins | LookADev

AI Systems Architect. I build systems that take repetitive admin off UK practices and small teams. No made-up answers, 7-day sprints. Free diagnosis below.

[Lucas Martins | LookADev @devlucasmartins]

A/B'd lcc vs the plugin most people already run, on tool calls.

lcc: -13.3/-39.2/-55.7%, fact recall 1.00/1.00/1.00.

Plugin: -70.5/-78.9/-83.4%, recall 0.00. Dropped the evidence every run.

Real sessions: -26.9% to -79.7%, 0 of 211 texts lost.

[Lucas Martins | LookADev @devlucasmartins]

lookacrawler contacts <url> pulls emails, phones, WhatsApp links and social profiles from a company site.
I use it in my own lead pipeline, locally.
https://t.co/4af4GsuEp5

[Lucas Martins | LookADev @devlucasmartins]

An agent cannot fix a stack trace it only knows from a summary.
lcc keeps the trace exactly as it was or drops it whole, and tells you which.
https://t.co/sfX2Oq23la

[Lucas Martins | LookADev @devlucasmartins]

An acknowledgement is not a booking. The form says received and reception never heard of it.

Test the path on a phone. Where does yours stop?

https://t.co/uxJEXJ56F7

[Lucas Martins | LookADev @devlucasmartins]

lookacrawler runs as an MCP server, so Claude Desktop or Cursor can extract, map and crawl a site without a hosted scraper or API key.
https://t.co/4af4GsvceD

[Lucas Martins | LookADev @devlucasmartins]

PetFoodFix scores a label against published FEDIAF and NRC reference values, and the criteria are public.
If a score looks wrong, you can check the rule that produced it.
https://t.co/GOfnomX6sk

[Lucas Martins | LookADev @devlucasmartins]

lcc ships as a CLI, an MCP server and a Claude Code plugin.
Same rule in all three: keep blocks verbatim or drop them, never paraphrase, write a receipt.
https://t.co/sfX2Oq23la

[Lucas Martins | LookADev @devlucasmartins]

Using Claude Code? This prints how much lcc would cut from one of your sessions. Offline, counts only:

pip install local-context-compiler
lcc measure-session ~/.claude/projects/<project>/<id>.jsonl --markdown

Where does it do worse than /compact?

[Lucas Martins | LookADev @devlucasmartins]

/compact rewrites your Claude Code session as a summary and loses the exact path or error you need later.

lcc keeps every message byte for byte and stashes old tool outputs the agent can recall.

266 of my sessions: median 23% fewer tokens, none lost.

https://t.co/sfX2Oq23la

[Lucas Martins | LookADev @devlucasmartins]

Correction to a post of mine from September: I said a local 1K model and a remote judge gave byte-identical output on a 44k token corpus. That run used a stub. With real weights the local model cut 0.5%, the judge 52%.

The repo changelog says so. I should have caught it.

[Lucas Martins | LookADev @devlucasmartins]

"Meat and animal derivatives" on a dog food label can mean almost anything.
PetFoodFix flags vague terms like that and separates them from named sources you can trust.
Free label search, UK brands:
https://t.co/gzOLRUPNyI

[Lucas Martins | LookADev @devlucasmartins]

Below a 25% gain, lcc tells you compaction is not worth it and leaves your context alone.
I would rather ship a tool that says no than one that always claims a win.
https://t.co/sfX2Oq2BaI

[Lucas Martins | LookADev @devlucasmartins]

A number without its sample size is a rumour with a decimal point.

Small sample, not claiming more. What is the smallest measurement you would still publish?

https://t.co/Dw3yEhdcEC

[Lucas Martins | LookADev @devlucasmartins]

Built https://t.co/GOfnomX6sk: type a dog food brand, get a plain-English read of the label.
Named proteins, vague meat terms, additives, daily portion, price check.
No brand sponsorship. The scoring criteria are published.
https://t.co/GOfnomX6sk

[Lucas Martins | LookADev @devlucasmartins]

A booking demo proves you can create an appointment. It does not prove you can change one.

When did your practice last test a cancellation?

https://t.co/1FRn4DP3aP

[Lucas Martins | LookADev @devlucasmartins]

@FishAudio DRAMA

[Lucas Martins | LookADev @devlucasmartins]

@BrenBuilds Just saw a person showing 134t/s using sol 6.1 in deepseek harness, can you make this test in other harness?

[Lucas Martins | LookADev @devlucasmartins]

@ThomasLisa97705 @MickMyers72 @ChrisBrooksDJ @AldiUK Most people don’t know how many hours they work each week or month. They calculated the hours incorrectly, and sometimes it is impossible to know. He probably only discovered this overpayment through this letter. Accusing someone of a crime like you did is itself a crime.

[Lucas Martins | LookADev @devlucasmartins]

An assistant with no exit is a liability. Agree the handoff before it goes live.

Routine, clinical or unclear. Which path is still undefined at your practice?

https://t.co/trwdZyYC4Q

[Lucas Martins | LookADev @devlucasmartins]

UK practice managers: send a test enquiry after closing. Can reception find its owner and next action in the morning?

https://t.co/yaf6r1tidl

[Lucas Martins | LookADev @devlucasmartins]

@thsottiaux Ask Tibo for reset button.
If more than 100k people click they receive a reset.