Logo
Picture

Parth Mudgal @phparth

CoFounder @100x.bot 🤖 Building AI automation tools for Chrome | Productivity enthusiast | Making browsers smarter with intelligent agents 100x.bot India Joined July 2011
  • Tweets
    607
  • Followers
    171
  • Following
    202
  • Likes
    152
phparth Profile Picture

Parth Mudgal @phparth

2 days ago

use “converging” prompts instead of “diverging” prompts for better results from llms. some examples where i noticed my prompt got into the “diverging” zone and the llm will start to drift. a rephrasing in specific direction can make it converging: eg A: diverging: “write about product strategy for this startup” converging: “given this startup sells X to Y users, write 3 product strategy options focused only on retention, with tradeoffs and a recommendation” eg B: diverging: “improve this landing page” converging: “rewrite only the hero section to make the value prop clearer for first-time founders, keeping the tone direct and non-salesy” the key difference is that diverging prompts open too many possible paths at once. converging prompts narrow the search space: define the audience, the goal, the constraint, the output shape, and what not to do. llms are much better when you give them a direction to converge toward, instead of asking them to explore everything and hoping they land in the right place.

phparth Profile Picture

Parth Mudgal @phparth

7 days ago

role/persona names themselves did not help much in maintaining real boundaries. calling something an auditor, planner, or theory keeper did not make it behave like one, you would still need to tell it what a "good auditor" does

phparth Profile Picture

Parth Mudgal @phparth

7 days ago

prompt instructions did not catch this reliably (each persona prompt kept growing, starting from about 400 lines to 800 lines). I saw the model classify a claim as unresolved, then still use that same claim as a constraint

phparth Profile Picture

Parth Mudgal @phparth

7 days ago

I tried a persona/FSM-heavy loop for complex SWEbenchpro tasks: surveyor, mapper, theory keeper, planner, auditor, worker, validator, reviewer. The bet was that specialization would make the system more reliable than a plain agent loop. So far, that bet has not held.

phparth Profile Picture

Parth Mudgal @phparth

a week ago

The wiki page of "Three wishes" need to be updated to include "using an LLM in 2026" the first wish is spent trying to get the model to understand the prompt, the second clarifying that no, you meant it literally, and the third undoing the consequences

phparth Profile Picture

Parth Mudgal @phparth

a week ago

@Karmedge somewhere a pareto principle among these 4

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

state machine + personas for the harness. each state will have an associated persona that will require a particular "handoff" document for the next state. the runtime fills in the "whats expected by possible next states" into the current state persona prompt

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

using swebench pro as the baseline. divided the tasks in terms of "ease" based on how many models solved the task. 6/6 models solved = easiest task, 0/6 = hardest.

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

working on my own agent harness, usign minimax 2.7 as the base model to make it work reliably, considering if it works reliably with minimax2.7 it would only be better when other stronger/larger models. named it #pragma

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

I feel a sense of lost opportunity when asking LLMs to do debugging on my behalf since everytime I debug an issue I would usually find a bunch of related issues and eventually a way to have "better" implemenation in fewer LOC. with LLMs its quite opposite. Those "related issues" become lost opportunity and 99% times the LLM resolves an issue by yet more code.

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

@Karmedge Weeks of prompting can save you hours of planning

Parth Mudgal retweeted
miguelsaavedra Profile Picture

Miguel Saavedra @miguelsaavedra

2 weeks ago

Bullish on open source models! Bullish on inference optimization!

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

Big money might soon be going towards skill writers to persuade them to include some "commercial" product in their skill. Would you want to sponsor a skill ?

phparth Profile Picture

Parth Mudgal @phparth

2 weeks ago

Insanity is doing the same thing over and over again and expecting different results. Or its LLMs in 2026.

phparth Profile Picture

Parth Mudgal @phparth

3 weeks ago

and its completely in #golang 100% native

phparth Profile Picture

Parth Mudgal @phparth

3 weeks ago

I made puppt, a cli to inspect and edit pptx files reliably. the biggest part of the code base is actually the renderer. I started with an approximation based renderer against a set of ~65 slides for testing. but that method quickly hit the limits since presentationml/drawingml are quite long and on top of that the actual renders are affected by the implementation level quirks (msppt vs apple notes vs libre vs google docs). so I start implementing the spec. i divided the spec into 12 broad milestones. this is promising and less stressful since with every change you are not fearing if the "mismatch pixel count is going to go down or up". the first mvp of puppt was taking about 7 second to render because my render pipeline was mvp. so no caching/optimizations. spent another day working on it. now it can render a slide <200 ms on average. check it out here github.com/artpar/puppt

phparth Profile Picture

Parth Mudgal @phparth

4 weeks ago

fool me once, shame on you; fool me twice, shame on me; fool me thrice, shame on llm

phparth Profile Picture

Parth Mudgal @phparth

a month ago

@tim_evans today i am excited to share

Trends for United States

Haaland
Norway
Carl
#loveislandseason8
Koulibaly
Dembele
Corey
Tino
Mikey
jungkook
Iraq
Ronnie
Odegaard
Alannah
Mendy
Sparkle
Keyon
Bottled
Steam Machine
Olise

You might like

Negocios digitales, internet, potencial humano, mentalidad, sistemas.
Cesar Lemus
@cesarlemus
267 Followers
Software engineer @Whop
Deep Pandya
@deeppandya91
32 Followers