pristine-skill
Health Warn
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 8 GitHub stars
Code Warn
- fs module — File system access in scripts/session-watch.js
Permissions Pass
- Permissions — No dangerous permissions requested
No AI report is available for this listing yet.
An agent skill enforcing the First-Time Principle: write every change in its final form — no patches, no what-comments, no leftovers, no local/deployed drift. Cross-platform: Claude Code, Codex, OpenCode.
Pristine

Write everything as if it were being written for the first time.
A cross-platform agent skill that enforces the First-Time Principle — the
discipline of landing every change in its final form: no patches, no
"what"-comments, no leftovers, no local/deployed drift.
Works with Claude Code, OpenAI Codex, OpenCode, and OpenClaw (open Agent
Skill format).
Why
In AI-assisted development, patches compound differently than in
human-only teams:
- Code is rewritten often, but docs and memory are the only bridge across
sessions — a stale note makes the next agent decide on a false premise. - Agents read the whole file, not a diff — every leftover special case
dilutes the signal of the real design. - A patch that "works" is the easiest thing to reproduce, because it
requires the least understanding. Each round of patching teaches the next
session to patch.
Pristine is the counterweight: each change, made as if for the first time,
keeps the system evolving toward clarity instead of entropy.
The four laws
| Law | Meaning |
|---|---|
| No patching | Fix the root cause, not the symptom. No if special cases, no copied-and-tweaked logic, no switches that route around a problem. Rewrite when the patch pile is worse than a clean rewrite. |
| Code explains itself | Names carry the "what". Comments only say "why": business rules, constraints, history traps. No explanatory comments, no commented-out code blocks. |
| No residue | No backups, drafts, dead code, or intermediate states. Superseded rules are updated in place, never appended as "as of…" notes. |
| Deployment parity | What runs in production is exactly what was reviewed locally. No remote-only quick fixes — the repo must keep representing reality. |
| Session cost | A session that grows without bound is the same entropy as a file that grows without bound. Around 15 turns, propose a reset. |
Install
Copy SKILL.md into your agent's skills directory:
# Claude Code
mkdir -p .claude/skills/pristine
cp SKILL.md .claude/skills/pristine/
# OpenAI Codex
mkdir -p .agents/skills/pristine
cp SKILL.md .agents/skills/pristine/
# OpenCode
mkdir -p .opencode/skills/pristine
cp SKILL.md .opencode/skills/pristine/
Usage
The skill activates automatically when you start implementation, refactor,
or bug-fix work, and whenever the "just patch it" urge appears. You can also
invoke it by name:
pristine— the formal name ("we work in pristine mode")virgin/first-time— aliases, recognized semantically via the skill
description
Session watch (optional companion)
scripts/session-watch.js enforces the 15-turn
rule mechanically: it counts user turns in the current session and prints a
reminder when the threshold is crossed. Wire it as a Claude Code Stop hook:
"hooks": { "Stop": [ { "matcher": "", "hooks": [ { "type": "command",
"command": "node /path/to/scripts/session-watch.js --threshold 15" } ] } ] }
The script reads the session transcript path from the hook's stdin — no
session discovery, no hardcoded paths, exit code always 0 (a reminder, not
a gate).
Examples
See examples/before-after.md for
patch-instinct vs. first-time-move side-by-side comparisons.
License
MIT © 2026
Reviews (0)
Sign in to leave a review.
Leave a reviewNo results found