A Claude Skill that front-loads clarifying questions, kills padded output, and locks corrections in as standing rules — so every response gets sharper instead of repeating the same mistakes.
Humans get better every time they make a mistake and understand something new. That ability isn't limited to humans. I wanted my Claude supercharged at learning and improving, and that is why I built this. Use the skill and Claude gets sharper with every step forward.
- Batches 5–10 clarifying questions with sensible defaults before starting substantial work, instead of trickling questions out one at a time across four turns.
- Produces dense, no-padding output — no hedges, no restating the question back, no explaining what you already know.
- Carries corrections forward as standing rules for the rest of the session, so a fix given once never has to be given twice.
- Calibrates depth to the task — a one-line fix does not get a five-question intake.
- Tracks improvement across the conversation, so round four is measurably better than round one rather than the same class of mistake in new clothes.
The protocol is domain-general, with dedicated playbooks for each:
| Area | Playbook |
|---|---|
| Coding | build order, interface-first, failure modes |
| Creative writing | voice, specificity, anti-slop |
| Documents and prose | structure, compression, register |
| Design and UI | taste, hierarchy, intentional choices |
| Analysis, research, decisions | evidence discipline, trade-off framing |
Plus cross-cutting sections: Thinking, Clarification Protocol, Improvement Rate, Depth Calibration, Efficiency, Verify, and a Question Bank.
Any task involving building, writing, coding, designing, analyzing or planning — especially short or open-ended requests like "build me X", "help me with Z", "make this better". Also when you tell Claude its output was generic, shallow, padded or off-target.
It deliberately skips lightweight Q&A and casual exchange. A factual lookup does not need a protocol.
Claude.ai / Claude Desktop — download getting-better.skill from the
latest release, then
upload it in Settings → Capabilities → Skills.
If no release is published yet, build the bundle from source — it produces the identical artefact:
git clone https://github.com/JeevaNadar1/Getting_Better.git
cd Getting_Better
make bundle # writes dist/getting-better.skillClaude Code — clone into your skills directory. The directory name must be
getting-better, matching the name: field in SKILL.md:
git clone https://github.com/JeevaNadar1/Getting_Better.git ~/.claude/skills/getting-betterClaude Projects — paste the contents of SKILL.md into project instructions. The file
is self-contained, so this works without modification.
- The question batch is capped at 10. It is an intake, not an interrogation.
- Defaults must be stated explicitly, not implied, so you can skip answering and still get a correct-by-default result.
- Nothing already answered or corrected earlier in the session gets re-asked.
- It scales down. Trivial tasks bypass the protocol entirely.
- Output favours substance over scaffolding — no filler headers, no repeated summaries.
getting-better/
├── SKILL.md the complete protocol, self-contained
├── README.md this file
├── CONTRIBUTING.md invariants, review bar, release procedure
├── CHANGELOG.md version history, Keep a Changelog format
├── CODE_OF_CONDUCT.md conduct standards and reporting route
├── LICENSE MIT
├── Makefile check / bundle / release-check / version
├── .gitattributes LF normalisation, language statistics
├── scripts/check_skill.py structural and drift validation
└── .github/
├── workflows/ci.yml validates, bundles, publishes tagged releases
├── PULL_REQUEST_TEMPLATE.md evidence checklist for protocol changes
└── ISSUE_TEMPLATE/ bug report and feature request forms
SKILL.md is currently a single self-contained file. It uploads to Claude as-is and reads
as documentation in the repository.
Read CONTRIBUTING.md first. The short version: make check must pass,
SKILL.md edits need before/after behavioural evidence in the PR, and because the file is
loaded in full on every trigger, additions are weighed against their permanent token cost.
Conduct standards are in CODE_OF_CONDUCT.md. Version history is in CHANGELOG.md.
MIT. See LICENSE.