Zero Drift

Zero Drift

No hallucination. No drift. No losing track.

Two rules. Active every response. No exceptions.

---

Rule 1: Named Response

Every reply starts with the user's name.

Purpose — hallucination detector. The name is the canary. When the model starts drifting or hallucinating, the name is the first thing to break: dropped, altered, or off-tone. The user watches for this. The instant it slips, the session is degrading — the user opens a fresh window, says "read the TASK.md and continue", and resumes clean. So: never drop the name, never alter it. A consistent name = a grounded session.

How to find the name

Run these in order, stop at the first hit:

1. git config user.name — use the result
2. Scan CLAUDE.md / AGENTS.md / GEMINI.md for a line like MY_NAME= or name: in a personal/user section
3. Check if the user introduced themselves earlier in this session
4. If nothing found → ask once:

   "What's your name? I'll use it throughout our session." Store the answer for the rest of the session.

Format

[Name]: [rest of response]

Example:

Breno: the login bug is in auth/middleware.ts:42 — token expiry uses `<` instead of `<=`.

Never skip the name. Never forget mid-session.

---

Rule 2: Living Task Document

Every specific task gets a TASK.md in the project root. Update it after every prompt that advances the task.

When to create

Create TASK.md when the user starts something specific:

• "Let's build X"
• "Fix this bug"
• "I want to create a Y"
• Any session where the work has a name and a goal

If a TASK.md already exists in the project root, read it before doing anything else in that task.

When NOT to auto-read

A new Claude instance does not read TASK.md automatically. The user must ask explicitly:

"Read the TASK.md and continue" / "Leia o TASK.md e continue"

Only then load and follow the document.

Document structure

# TASK: [Task Name]
> Created: YYYY-MM-DD | Updated: YYYY-MM-DD HH:MM

## Goal
[What we're building/fixing — 2-3 sentences max]

## Plan
- [x] Completed step
- [ ] Pending step

## Log
### YYYY-MM-DD
- Added retry to fetchUser() — `npm test auth` → 12 passed, exit 0
- Fixed null deref in parse() — `cargo test parse` → ok. 3 passed, exit 0
- Ran migration 0007 — `psql -f 0007.sql` → ALTER TABLE, exit 0

## Unverified / Pending
- Refactored cache layer — NOT tested yet, no proof
- Believe the timeout bug is fixed — could not reproduce, unconfirmed

## Errors & Fixes
| Error | Cause | Fix | Evidence |
|-------|-------|-----|----------|
| `TypeError: X` | Missing null check | Added guard at line 42 | `npm test` → exit 0 |

## Current State
[1-2 paragraphs for a fresh Claude instance: what was built, what works,
what's broken, what's next. Write this last. Keep it current.
Only claim something works if its proof is in the Log. No proof → say
"implemented, not verified", not "done".]

Update rules

• Update after every prompt that does something meaningful
• Current State always reflects right now — rewrite it, don't append
• Keep the file readable — no noise, no filler

Evidence rule (the anti-drift core)

The Log is a record, not a diary. The model does not get to assert what it did — it shows proof. When context degrades, the model pulls toward fluency before truth: it writes "fixed X" because that reads well, not because X is fixed. The fix is to ban the assertion and demand the evidence.

• Log = evidence only. Every line claiming "did / fixed / works" carries raw proof inline: the command run + its output + exit code. Test result, build output, a reproduced behavior. No proof, no Log entry.
• No proof → Unverified / Pending. Work done but not yet proven goes here, marked plainly: "NOT tested", "unconfirmed", "could not reproduce". Never launder it into the Log.
• Current State is derived, not declared. It can only claim something works if that proof sits in the Log. Otherwise the strongest allowed phrasing is "implemented, not verified". A fresh instance trusts this file because every "works" is backed — it reads a record, not a story.
• Logged a fix? Log both sides with proof: what broke (error + how it surfaced) and the evidence it's now fixed (command + exit code).

Handoff protocol

When starting a new session on an ongoing task, the user will say something like:

• "Read the TASK.md and continue"
• "Leia o TASK.md e continue"
• "Continue from where we left off — check TASK.md"

On that trigger:

1. Read Current State first
2. Read Log (most recent entries)
3. Read Plan (check vs uncheck)
4. Confirm understanding: "Got it. We're at [X], next is [Y]. Continuing."
5. Do not redo completed steps

---

Activation

Recommended — install as a plugin so a SessionStart hook activates these rules in every session automatically:

/plugin marketplace add obrenoalvim/zero-drift
/plugin install zero-drift@zero-drift

Manual fallback — paste this into ~/.claude/CLAUDE.md:

# Zero Drift Skill
Always follow the Zero Drift rules from: https://github.com/obrenoalvim/zero-drift

Rules:
1. Start every response with the user's name (detect from git config, CLAUDE.md, or ask)
2. For every specific task, maintain and update TASK.md in the project root

Or paste the full SKILL.md content directly into your CLAUDE.md.

---

Quick Reference

Situation	Action
First response, name unknown	Ask for name, then use it
Name found in git config	Use it immediately
New task starts	Create TASK.md
Task already has TASK.md	Read it first
New session, same task	Wait for explicit "read TASK.md"
After every meaningful prompt	Update TASK.md log + Current State
Claiming "fixed / works"	Only in Log with raw proof (command + output + exit code)
Done but not tested	Goes in Unverified / Pending, never the Log