iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
In Daniel Miessler’s LifeOS workflow, Jev evaluates a prompt together with the last part of the preceding reply, answers 18 structured questions, and passes those answers to two learned models that select a model lane and reasoning effort. Jev does not write the task response, and the resulting route is advice—not an automatic dispatch. Daniel’s explicit instructions take precedence.
What Jev looks at before choosing
Jev receives the current prompt and, when available, the final 800 characters of the preceding reply. This context matters because a short follow-up may depend on what was just proposed or answered; the router does not have to interpret it in isolation. About three quarters of the prompts in the reported evaluation had a previous reply attached.
Jev responds with probabilities or structured choices rather than generated prose. It answers 18 typed questions about the request, its difficulty and risk, and its relationship to the preceding reply. Three additional prompt facts—whether it contains words requesting depth, its length, and whether a previous reply exists—also feed the selection process.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11The 18 questions
The earlier design used nine questions about the work itself. The mature design kept those and added five about reasoning effort and four about conversational context:
#1 Best Overall
- Work and task shape: the nine questions examine what the prompt asks for, its risk and breadth, its potential for automation, whether the approach is settled, and how much judgment or reasoning it needs.
- Reasoning effort: five questions consider breadth versus depth, the cost of subtle errors, interacting considerations, whether speed matters once the task is understood, and whether the user explicitly asks for depth.
- Conversation context: four questions assess whether the user approves a proposal, acknowledges it, corrects or pushes back on it, or is making a new request.
Two small learned models map Jev’s answers and the three prompt facts to a model lane and an effort level. Jev’s answers inform the choice; they do not directly name the final model or effort.
How Glance turns the assessment into a route
Glance is the surrounding judgment and control layer. It handles thresholds for callers, a daily budget, and a decision ledger, while Jev supplies the structured assessment. New callers begin in shadow mode. Moving to enforcement requires a recorded agreement rate, date, and Jev model. The LifeOS article says the router itself remains in shadow, so its recommendations should not be read as evidence that it is automatically sending work to another model.
The routing design developed in stages. An initial Jev version selected among seven lanes and matched a prior Astra classifier on 57% of 662 prompts. The next version broke the decision into nine yes-or-no questions and first mapped those answers by hand, then with a small learned model. Synthetic prompts produced misleading confidence, so the later evaluation instead used real interactive prompts sampled from Daniel’s transcripts.
Free tools Windows power users keep installed
One-click scans. No signup required.
The evaluation used 1,000 prompts. Automated jobs, hook output, pasted notifications, and messages from other agents were excluded; the prompts remained private on Daniel’s machine. Testing held out data by conversation, rather than treating prompts from the same session as independent examples. The source does not publish the prompt-level labels, so the reported result cannot be independently reproduced from the article alone.
What the reported accuracy figures mean
Daniel Miessler’s LifeOS team reported these results in 2026. Lane agreement compares a selected lane with majority-vote labels from three model labelers; effort match compares the chosen effort with those reference labels.
| Measure | Reported result | How to interpret it |
|---|---|---|
| Glance lane agreement | 90.1% | Agreement with the majority-vote lane labels on the 1,000-prompt evaluation. |
| Opus classifier lane agreement | 75.2% | Baseline result on the same comparison. |
| Always-inline lane agreement | 83.9% | Baseline result on the same comparison. |
| Glance effort match | 70.0% | Agreement with the majority-vote effort labels. |
| Agreement among the three labelers | 79.4% | The labelers did not always agree with one another. |
| Reported routing time | About 0.3 seconds per prompt for Glance; about 3.3 seconds for the Opus classifier | Implementation timings reported by the author, not a general benchmark. |
The 90.1% lane figure is not a measure of whether the chosen model completed a task successfully. It measures agreement with labels produced by this particular process on a private dataset. The labelers’ 79.4% agreement also shows that the reference itself was not perfectly consistent. These results do not establish performance for other users, prompts, or routing systems, and they do not demonstrate cost savings.
Rules and exceptions that affect a choice
When a task is considered settled
The article’s practical distinction between Opus and Sol is whether a pass/fail check can be written before work begins. If the work has a clear test in advance, it is treated as settled and Sol can handle it. This is a rule described for this workflow, not a universal definition of which tasks require a more capable model.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteDepth requests and maximum effort
A request to “think deeply,” or similar wording, forces Opus at xhigh effort. Max-level work is also routed to Opus at xhigh; Fable is used for second opinions. The Opus classifier is used if Jev times out or returns an incomplete answer, and it also selects thinking skills for depth prompts.
Prompts that skip ordinary routing
- Acknowledgements and slash commands skip routing.
- Prompts that look like credentials skip routing.
- Email addresses and phone numbers are redacted before Jev or Opus sees the prompt.
Advice, not automatic dispatch
Kai, identified in the LifeOS article as Daniel Miessler’s AI assistant, describes the pick as advice: it does not dispatch work by itself, and Daniel’s explicit instructions always win. That boundary is important: the workflow can recommend a lane and effort without overriding a user’s stated direction.
A separate Jev Codex router implementation illustrates different configurable safeguards: it builds eligible routes from models’ supported effort levels, classifies task capability and request kind, and adjusts choices for routing preferences, effort ceilings, and usage policy. That is a separate implementation example; it is not evidence that the LifeOS router uses the same code or controls.
What the design is meant to improve
The central design choice is to make model selection a structured judgment rather than a single direct classification of the prompt. Including the preceding reply gives the system conversational context; separating Jev’s assessment from the two learned mappers creates distinct stages for evaluating the request and selecting a route. Glance then provides thresholds and logging around that process.
Recommended Free Tools
The reported figures support a limited conclusion: on the team’s labeled, session-held-out set of 1,000 private prompts, Glance agreed with the reference lane labels more often than the two reported baselines. They do not establish that every selected model is optimal, that the task outcome is better, or that the same rates would hold in another person’s workflow.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
