Conversation Drift Audit and Reset
Find where a conversation went off-goal and get a minimal recovery plan that preserves useful work.
Before & after
Same prompt, same agent — the only difference is the skill.
Sure! Let me summarize where we've gotten to. We started by looking at your authentication issue, then explored a few different approaches. We tried adjusting the session middleware, then moved on to debugging the token refresh logic, and most recently we've been working on the Redis connection pooling. We've made good progress on several fronts. The session handling is cleaner now, and we've identified some improvements to the caching layer. Shall we keep going with the connection pooling, or would you like to focus on something else?
20K FEET WHAT YOU ASKED FOR (message 1) "Users get logged out randomly after about 15 minutes." WHAT WE ACTUALLY DID 0:00 Reproduced it. Confirmed 15-min logout. ✓ on track 0:20 Found session middleware refreshing on every request 0:35 Refactored the middleware ← first drift 1:10 Refactor surfaced a token-refresh race, fixed that 1:40 Fixing the race led to Redis connection pooling 2:00 Currently tuning pool size WHERE THEY DIVERGED At 0:35. You had the diagnosis — the middleware was refreshing every request instead of on expiry. That was a one-line fix. The refactor was optional work that generated its own follow-on work, and each step after it was a legitimate response to a problem the previous step created. IS THE ORIGINAL BUG FIXED? Unverified. We have not re-tested the 15-minute logout since 0:35. CORRECTED PLAN 1. Stash the pooling work. It's real, it's just not this. 2. Re-test the original repro. It may already be fixed. 3. If fixed: ship the one-line change alone, file the rest as follow-ups. 4. If not: resume from the 0:20 diagnosis, not from here. Worth naming: I'd have happily kept tuning the pool. Each step looked reasonable from the step before it. That's the failure mode this exists to catch.
About this skill
“Why are we rebuilding the pipeline? I asked for one small fix.” When a long conversation turns a side problem into the main project, this skill reconstructs the intended outcome, identifies the supported point of divergence, and gives you the shortest route back.
It separates legitimate changes you requested from scope Claude introduced, so useful discoveries survive without becoming excuses to keep going.
What it does
- Restates the current goal using your original request and later explicit changes.
- Locates the first supported scope expansion with a short quote or message reference.
- Tests whether side work is a real prerequisite or an optional improvement.
- Recommends continuing, narrowing, parking a tangent, or restarting with a clean brief.
- Separates work to keep from work to stop, with concrete conditions for revisiting it.
- Recaps visible effort without inventing token counts, costs, or drift percentages.
What you get
A concise drift verdict, evidence, and a recovery plan with a clear finish line. For a hard reset, you also get a paste-ready restart brief containing the goal, constraints, useful decisions, and next action.
Best for
People using Claude for long coding, research, writing, or planning sessions who suspect the conversation is solving a different problem from the one they brought.
SKILL.md preview
---
name: conversation-drift-audit-and-reset
description: Audit a conversation against the user's authorized goal, trace the first supported scope divergence, and produce a minimal recovery plan with keep and stop decisions. Use when the user says "step back", "we're in a rabbit hole", "why are we doing this?", "this conversation keeps going sideways", "I only asked for one small change", "reset this thread", or "are we still solving the original problem?".
---
# Conversation Drift Audit and Reset
… (sign up to view the full skill)More productivity skills
View all Productivity skills →Apple-Style UI Engineer
Use when building any new UI component, screen, or page in React + Tailwind from a plain-text description, and the result must look like App
Zapier, Make and n8n Automation Recipe Builder
Designs no-code automations that run unattended: exact trigger, mapped fields, dedupe, error alerts, approvals, and monthly task cost.
AI Productivity System
A complete AI-driven system for planning your day, triaging tasks, and shipping more without burning out.
Daily Standup Writer from Git and Tickets
Turns your commits, PRs, reviews, and ticket moves into a 15-second standup post that never says shipped when it only merged.