think-better

Think Better

Structured thinking for your AI coding assistant: decide with frameworks, solve with root causes, code with evidence. One CLI installs three skills and their slash commands so your AI decides with real frameworks, solves problems with a proven method, and changes code with evidence instead of guesses.

3 skills · 326 knowledge records · 20 slash commands · 4 AI tools

Works with Claude Code · GitHub Copilot · Antigravity · OpenCode

Website · User Guide · Quick Reference · Examples · Changelog

Install

Pick one:

# Homebrew (macOS / Linux)
brew tap HoangTheQuyen/think-better https://github.com/HoangTheQuyen/think-better && brew install think-better

# Scoop (Windows)
scoop bucket add think-better https://github.com/HoangTheQuyen/think-better; scoop install think-better

# Install script (macOS / Linux) — installs to ~/.local/bin and verifies the SHA-256 checksum
curl -fsSL https://raw.githubusercontent.com/HoangTheQuyen/think-better/main/install.sh | sh

# Install script (Windows PowerShell) — installs to %LOCALAPPDATA%\think-better
irm https://raw.githubusercontent.com/HoangTheQuyen/think-better/main/install.ps1 | iex

# Go 1.25+
go install github.com/HoangTheQuyen/think-better/cmd/think-better@latest

Then, inside your project, install the skills for your AI tool:

Every target gets all three skills and all 20 slash commands (/solve*, /decide*, /code*). Add --global to install once for every project (for GitHub Copilot this installs the skills, in ~/.copilot/skills/, but not the slash commands), --skill code-solving to install a single skill with its commands, or --exclude-command code.perf to leave out a slash command you do not want. The skills need Python 3 (standard library only); run think-better check to verify.

After upgrading the binary, run think-better update to refresh every install; files you edited are kept (see Updating).

# Pin a release and pick the directory (no sudo needed)
curl -fsSL https://raw.githubusercontent.com/HoangTheQuyen/think-better/main/install.sh \
  | THINK_BETTER_VERSION=v1.5.0 INSTALL_DIR="$HOME/bin" sh

# Nix
nix run github:HoangTheQuyen/think-better -- init --ai claude

# From source
git clone https://github.com/HoangTheQuyen/think-better && cd think-better && make build

Manual download: grab a binary from Releases and check it against checksums.txt.

30 seconds per skill

Type a slash command, or just describe the problem: the AI also picks the skill from its description and trigger phrases (see How a skill is picked). The samples below are real script output for these requests, shortened.

/decide: choose between options

/decide Postgres vs MongoDB vs DynamoDB for our order service

→ Decision type: Multi-Option Selection · Options: Postgres | MongoDB | DynamoDB
→ Framework: Weighted Criteria Matrix (criteria and weights before scoring)
→ Criteria (Tech Stack / Framework Choice): team expertise 25, ecosystem 20, performance 20,
  maintainability 20, cost and licensing 15
→ Bias warnings: Anchoring Effect, Availability Heuristic, Confirmation Bias, each with a remedy
→ A scoring command, a checklist, then next steps: /decide.deep, /decide.exec

/solve: get to the root of a business or product problem

/solve Signups dropped 15% after the pricing change

→ Type: Diagnostic · Context: Business Performance
→ Process: Define precisely → Profitability tree → Pareto prioritize → Hypothesis-driven analysis
→ Decomposition: Fishbone (Ishikawa) · Prioritize: Pareto Analysis (80/20) · Analysis: Root Cause Analysis (5 Whys)
→ Bias warnings: Confirmation Bias, Narrative Fallacy, Availability Heuristic
→ Communicate with the Pyramid Principle; next steps: /solve.deep, /solve.exec, /decide

/code: change code in 7 gated steps

/code.debug Checkout throws after yesterday's deploy when paying with a gift card:
            TypeError: Cannot read properties of undefined (reading 'id')
                at orderTotal (src/checkout/total.js:10:33)
                at payHandler (src/checkout/handler.js:4:17)

→ Context from the project: src/checkout/total.js:10   const discount = cart.coupon.id ? ...
                            src/checkout/handler.js:4  const total = orderTotal(req.session.cart);
→ Known error: TypeError: Cannot read properties of undefined/null (likely causes, what to
  check first, the root-cause fix)
→ Steps: Define (a failing test) → Decompose → Prioritize → Plan → Execute → Verify → Communicate
→ Project checks found: npm run test, npm run lint, npm run build

Full walk-throughs: examples/.

The three skills

/decide — make a choice · make-decision · 63 records

/solve — solve a problem · problem-solving-pro · 114 records

For bugs and other code changes use /code.

/code — change code · code-solving · 149 records

Slash commands

Depth: .quick is a fast scan, the plain command is the default, .deep adds alternatives and pitfalls for high-stakes work, .exec adds a summary for leadership. /code auto-detects the task type; /code.deep is the same with more techniques and the full review checklist.

Add "save step-by-step" to any request to get a Markdown workspace with one file per step (solving-plans/, decision-plans/ or coding-plans/). Saving again keeps the files you already filled in.

Why not just prompt?

You can ask any AI "should we use Postgres or MongoDB?" and get a tidy pros-and-cons list. The skills add what a prompt alone does not:

• Structured steps. Every request runs through a fixed method (7 steps for problems and code, a decision plan for choices), so the AI does not jump from symptom to fix.
• Gates. /code does not move on without evidence: a failing test before the fix, the project's own test/lint/build output after it.
• A knowledge base. 326 records (frameworks, cognitive biases with remedies, criteria templates, 44 known error messages, ...) are searched locally and put into the answer, so the advice is specific and the same request gets the same method every time.
• Saved, resumable workspaces. Say "save step-by-step" and the work is written to Markdown files in your project; /solve.resume, /decide.resume or /code.resume continues at the first open step in a later session.

How a skill is picked

Slash commands work in every tool and set the depth for you. Copilot installs made by v1.4.0 or earlier put the skills in .github/prompts/<skill>/, where Copilot does not load them on its own; think-better update moves them to .github/skills/.

The knowledge base and the trigger phrases are in English, so write requests in English for the best skill match.

How it works

You ── "Revenue dropped 20%"  or  /solve.deep …  or  /code.debug …
          │
          ▼
  AI tool (Claude Code · GitHub Copilot · Antigravity · OpenCode)
   ├─ picks the skill: SKILL.md trigger phrases, or the slash command
   └─ runs the skill's script:  python3 <skills dir>/<skill>/scripts/search.py --stdin --plan
          │
          ▼
  Skill engine (local, Python 3 standard library)
   ├─ BM25 search over 326 knowledge records (CSV files shipped with the skill)
   ├─ classify: problem type · decision type · coding task type
   ├─ /code only: read the project (stack-trace frames, symbols, git log, diff, test commands)
   └─ build the plan: framework · steps and gates · bias warnings · checklist
          │
          ▼
  A structured answer + next-step commands (optionally saved as a step-by-step workspace)

The CLI only copies Markdown, CSV and Python files into your project (or home directory); the scripts run locally with no network calls, accounts or API keys.

CLI