Mumbai > Bengaluru…..
for the diverse set of people you see everyday around
For the weather Blr>>>
1
5
173
Shrish Dwivedi retweeted
JEV is INSANE
It read 6,030 ChatGPT answers to buyer questions about 1,005 YC companies and checked every company name in every answer.
37,636 verdicts.
9.1M tokens.
40 seconds.
$0.38.
Grading the same 6,030 answers with Claude Sonnet cost me $36 last month.
71% of these startups were never named once, even when the buyer asked ChatGPT for exactly what they build.
Scan your site and see if AI mentions you. Link in the first comment 👇
Shrish Dwivedi retweeted
🚨 ¿ENTIENDES LO QUE SIGNIFICA QUE EL 35% DEL FORTUNE 500 YA USE ESTO?
Ryanair, Volkswagen, Bayer, Roche.
Mismo problema que tienes tú, a otra escala.
El conocimiento vive en once documentos que nadie abre, y el agente acaba inventándose la respuesta.
Trail lo convierte en reglas que cualquier agente puede seguir.
→ Subes un SOP y te construye el agente
→ Cada regla enlaza al texto exacto
→ Nada se activa sin tu aprobación
→ Funciona con Claude, ChatGPT y Copilot
Deja de adivinar.
Guárdalo 🔖
🤝 Paid partnership
Shrish Dwivedi retweeted
create your Trail Brain here: trailhq.com
Shrish Dwivedi retweeted
Count your skill files. Now count how many contradict each other.
That's what it costs when every session starts from scratch.
You write the rule down. The agent reads it, does the work, session ends. Next session, fresh context, same file. It learns nothing it didn't already know. Makes the same mistake you corrected last week.
Skills gave agents instructions. They never gave them memory.
Trail turns one-off skill files into brains that compound. The key: every rule is stored separately, scoped to the decision it governs, and kept current by Claude Code hooks while you work.
→ It learns from what you actually do, not from what you remembered to write down
→ Claude Code hooks keep it current, so you are not maintaining it by hand
→ It knows when a rule became true, separately from when it learned it
→ Every rule carries an owner
Skills aren't dead.
They're evolving into Trail brains that improve themselves.
🤝 Paid partnership
Shrish Dwivedi retweeted
one of the repos going around SF developer circles right now just crossed 5K stars
and it made me realize something about PR reviews:
we spend way too much time reviewing the diff
and not enough time asking what the diff can actually break
every PR has 3 things that are easy to miss
a test sitting next to your change doesn't mean your change is tested
it means a test exists
those are different things
a 10-line PR can also break the exact same 8 files downstream as a 10,000-line PR
so "keep your PRs small" isn't really the answer
size was never the problem
reach was
the tool is Graft.
it traces the actual dependency graph from what you changed
not just the files in the diff
it shows the exact line proving a dependency
flags downstream tests that weren't touched
and suggests reviewers based on the code your change actually reaches
that's the part I find interesting.
it's answering:
"what can this change actually affect?"
instead of:
"what files changed?"
dropping it in the replies.
🤝 Paid partnership
Shrish Dwivedi retweeted
🚨 6 repos de AGENTES de IA que están explotando en GitHub
1. Graft → Claude Code más barato y más rápido
2. Agency Agents → 232 subagentes en 16 áreas
3. Codebase Memory → tu código en un knowledge graph (menos tokens)
4. OpenMontage → de agente de coding a estudio de vídeo
5. Agent-Reach → ojos en X, Reddit, YouTube y más (sin APIs de pago)
6. Orca → varios agentes de coding en paralelo
De un solo agente a una flota completa.
Guárdalo 🔖
LE DAS 10 LÍNEAS A ALGUIEN Y ENCUENTRA 10 FALLOS.
LE DAS 10.000 Y LO APRUEBA SIN MIRAR
Todo el mundo lo ha hecho.
El PR pequeño se lee línea a línea. El grande se hojea y se aprueba, porque quién tiene una hora para rastrear 10.000 líneas.
El problema es que el tamaño no tiene nada que ver con el riesgo.
Un cambio de 10 líneas puede tocar los mismos 8 archivos que uno de 10.000. El PR minúsculo que aprobaste sin mirar puede ser el peligroso.
Tú ya lo sabes. Por eso revisar tarda tanto: no estás leyendo el diff, lo estás rastreando. Quién llama a esto. Qué se rompe si cambia. Si hay un test que lo cubra.
Graft hace ese rastreo por ti. Cada PR llega con un mapa: qué cambió, qué depende de eso, si algún test lo cubre.
→ Menos tiempo de revisión
→ Sin LLM, sin clave de api
→ Nada sale de tu repo
→ Open source
Se llama Graft, os dejo la repo abajo ⬇️
os dejo la repo por aquí: github.com/trailhq/Graft
Shrish Dwivedi retweeted
I found an open source tool that stops coding agents from relearning your codebase every session.
A new engineer joins your team and learns the codebase once. A coding agent learns it every single session, greps around, follows imports, builds a picture of how everything connects, and then throws that picture away the moment the chat ends. The next session starts blind again.
Graft ends the re-onboarding. It reads your repo and writes what it learns into a folder of linked markdown files, one node per system or concept, each explaining in plain English what that part of the code does and how it connects to the rest.
The clever part is what it doesn't use. There are no embeddings, no vector database, and no index server. The map is just files in your repo, so your agent opens it the way it opens any other file, and you can even commit it so a stale map shows up as a diff in code review.
It wires into Claude Code, Cursor, Codex, and Gemini with one command. The maintainers benchmarked it on SWE-bench and report the agent solving 66% of issues with the map against 54% without, while running up to 4x cheaper and 3x faster.
The repo sits at 3,000 stars, the whole thing is MIT licensed, and the link is in the reply.
Shrish Dwivedi retweeted
The biggest problem with Claude Code has just been solved.
Everything your agent learns from your code gets wiped out when you close the session.
And tomorrow it learns it all over again. And your teammate learns it from scratch again with their own agent. And again. And again.
Nobody on your team is saving that work anywhere.
This turns it into files:
→ Generates a repo map in markdown
→ Saves to your project
→ Updates only when the code changes
→ You share it via git and it benefits the whole team
→ Runs locally, no accounts or keys needed
A new hire reads the docs once. Your agent hasn't read anything in months.
It's called Graft. Open source, MIT license.
Link: github.com/NanoNets/Graft
Shrish Dwivedi retweeted
Aquí tenéis el repo: github.com/NanoNets/Graft
Shrish Dwivedi retweeted
HAY UNA FORMA DE GASTAR UN 42% MENOS DE TOKENS SIN BAJAR DE MODELO NI TOCAR NADA
Sin mapa del repo, tu agente entra en cada sesión con la memoria en blanco. Abre archivo, no era. Abre otro, cablea a ojo. Lo que averigua se pierde al cerrar.
Con el mapa delante, entra sabiendo qué sistema toca y qué se rompe si lo mueve.
Lo midieron con 162 pruebas controladas, mismo agente y mismas herramientas:
→ 42% menos tokens
→ 46% menos llamadas a herramientas
→ 60% menos tiempo por tarea
→ 54% de acierto en frío contra 66% con el mapa
Nadie cambió de modelo. Nadie pagó más. Nadie tocó un prompt.
Un becario que lleva un año en la empresa no es más listo que el que entró ayer. Solo sabe dónde está todo.
Se llama Graft. Gratis, MIT, y funciona con Claude Code, Cursor, Codex y Gemini. Tienes el repo abajo.
🔖 Guárdalo, te servirá.
Shrish Dwivedi retweeted
Votre agent lit tout le codebase à chaque prompt, et vous payez pour ça. Graft, open source, construit un graphe du projet et n'envoie que les nœuds utiles à Claude Code, Cursor ou Codex.
Sur SWE-bench : 42% de tokens en moins, 60% de temps gagné, +12 points de correctness. ⬇️


















