Same agent.Better habits.

stop coding. think first. finish one thing. then next thing.

AI coding agents rarely fail because they don't know the answer. They fail because they start before asking, move on before checking their work, and ship UI they never looked at. BetterThink puts six rules directly into the agent's prompt.

View on GitHub

The problem

Capable, just undisciplined

Same model, same task, both columns. The only difference is what it was told to do first.

Default agent behavior

  • Starts coding immediately
  • Assumes scope
  • Moves on before the current task is verified
  • Logs every tiny refactor into CLAUDE.md
  • Ships UI without opening a mobile viewport

BetterThink behavior

  • Asks until 95% confident
  • Scope confirmed, never guessed
  • One task at a time, verified before the next
  • Memory file updated only on meaningful changes
  • 375 / 768 / 1440 checked every time
Assumptions before codingmany
none
Tasks finished before nextsome
100%
Memory file bloathigh
none
UI viewport coverage1/3
3/3

What the six rules are measured against.

The skill

Six gates every task passes through

The agent reads these before it writes a line of code, so it can't pretend it didn't know.

01

Clarify before coding

threshold 95%

Ask until you're at least 95% sure what the user wants. Caught yourself filling a gap with a guess? That guess is a question you haven't asked yet.

FixesSilent assumptions about scope, stack and edge cases

02

One to-do at a time

verify → advance

Run it, read the diff, check the behavior. Only then move to the next item. No hopping away with an "I'll fix it later" attached.

FixesHalf-finished work left behind by task hopping

03

Delegate mechanical work

claude-haiku-4-5

Search, file reads, linters, screenshots and simple checks go to a Haiku subagent. Architecture, debugging and security review stay on the main model, where the judgment is.

FixesA main thread clogged with grep results and lint output

04

Keep memory files lean

150-200 lines

Update it for a new dependency, a new script, a new top-level folder, a changed run command, or a new convention. Nothing else. Cap it at 150 to 200 lines and prune before you add.

FixesA CLAUDE.md that grew into a changelog

05

Screenshot-verify all UI

375 / 768 / 1440

Run it, screenshot at 375, 768 and 1440, look at each layout, fix, repeat. "It compiles" is not a UI check.

FixesMobile bugs that only surface after release

06

Deep docs, loaded on demand

index, not archive

Write architecture.md or decisions.md as long as you want. Then link each one from CLAUDE.md with a single line saying when to open it.

FixesContext burned on documentation the task never needed

Install

One file, one copy command

Pick your agent. The skill turns itself on once it spots a coding task, so there is nothing else to configure.

Step 1 · Clone the repo

bash
$ git clone https://github.com/RokyYTR2/BetterThink.git$ cd BetterThink

No build step, no runtime, no telemetry. It's a prompt file.

Step 2

claude/
$ mkdir -p ~/.claude/skills/betterthink$ cp claude/SKILL.md ~/.claude/skills/betterthink/SKILL.md

Support

Five agents, one set of rules

Every folder holds the same six rules, written in the format that agent actually reads.

Claude Codeclaude/CLAUDE.md / AGENTS.md
OpenAI Codexcodex/AGENTS.md
Gemini CLIgemini/GEMINI.md / AGENT.md
Cursorcursor/.cursor/rules/*.mdc
GitHub Copilotcopilot/.github/copilot-instructions.md

Why

Discipline is the missing part, not capability

The agent knew how to fix the bug. It just moved on before checking whether the fix worked. It knew you wanted a mobile-friendly layout, but never opened a 375px viewport. It knew the memory file was supposed to stay lean, and logged every minor refactor anyway.

BetterThink is a set of guardrails written straight into the skill prompt. Nothing clever, nothing hidden. Six rules the agent reads before every task.

Get BetterThinkRead the full skill