Source: raw/How_to_Use_Fable_In_Every_Aspect_Of_Your_Life.md (youtube.com/watch?v=aPUvTP5NIUI)
A single-practitioner technique for using a screenshot loop, a cheap model, and a periodic Fable review to find and fix small workflow inefficiencies. Framed as a self-optimization system rather than a coding or content technique — distinct from Computer Use’s screen-driving pattern because this loop never acts, it only observes and reports.
Key Takeaways
- The loop, end to end. A screen-watcher script takes a screenshot roughly every 5 seconds while you work, logs it to a daily folder/database, and passes it to a cheap model to summarize (~20 sampled frames per pass, vision only where text alone can’t tell what happened — the creator estimates the raw log at roughly 8,640 lines of metadata, “basically free” because it’s text). The cheap model appends to a running observations ledger; once a specific low-value behavior shows up 3-4+ times, it’s confirmed as a pattern rather than a one-off. On a cadence you choose (daily or weekly), you ask Fable directly: “how can I economize my workflow?” — Fable is the expensive reviewer called in occasionally, not the model running the constant background summarization. This is the same cheap-workhorse/expensive-reviewer split documented in Cost & Intelligence Levers for Agent Workflows, applied to personal-productivity coaching instead of coding.
- Concrete outputs the creator got from under an hour of real screen time: (1) a suggestion to replace a polling habit (repeatedly re-checking the same pages through the day) with one high-quality morning digest bundling the same information; (2) a hotkey-based idea-capture system — a previously 5-10-minutes/day manual routine turned into a single Raycast script command with a Linear API passthrough (press a hotkey, dictate a line, back to work); (3) a suggestion to consolidate several views into one cross-team project-manager view.
- A concrete cost/latency win, not just a workflow win. The system flagged a paid voice-transcription tool (~$15/month, ~250ms latency per use) as a repeat-use pattern (hundreds to thousands of uses/day) and proposed a free local alternative that cut latency to ~50ms — the savings compound because of use-count, not because any single instance mattered.
- The “aggregate small waste” thesis. The creator’s framing: five-, ten-, fifteen-second frictions (an extra mouse trip, a menu instead of a hotkey) look trivial individually but compound across hundreds of thousands to millions of knowledge workers into meaningful lost output — explicitly presented as a claim about aggregate economic impact, not a measured figure. ^[This is the creator’s own framing/argument, not a cited statistic — treat as an argument, not data.]
- Scales down to a background/overnight job. Because the screenshot folder fills continuously regardless of whether you’re watching, the review pass itself can run on a schedule (e.g. once every 24 hours, or specifically before sleep to use up an otherwise-wasted daily session-limit window) rather than needing to be triggered manually.
- The creator’s own stated results: roughly 30+ minutes recovered per day within the first few days of running the system, with an expectation of eventually approaching 2-3 hours/day — self-reported, from someone who describes themselves as an already-optimized knowledge worker, so likely not representative of a typical user’s gains. ^[Single self-reported, non-independently-verified result.]
- Onward idea (unimplemented at time of recording): extending the same passive-observation loop from a computer screen to a physical room via a camera capturing stills every 5-10 seconds over several days, to catch physical/postural inefficiencies (screen height, lighting, chair position) the same way the desktop version catches digital ones — explicitly described as speculative, not yet built.
How It Differs From Computer Use
This is a passive observation-and-report loop, not an acting agent. Compare to Hermes + Computer Use, where the screen-recording loop exists so the agent can drive the screen (click, type, navigate) — there the screenshot feed is input to action. Here the screenshot feed is input to a workflow-efficiency report; the model never touches the mouse or keyboard, it only counts repeat behaviors and proposes fixes for a human to adopt.
Implementation
Tool/Service: any vision-capable model for the constant cheap summarization pass (the creator uses “Fable or another AI” generically for the watcher role, reserving Fable 5 itself for the periodic strategic review) plus a local screenshot-capture script running on an interval.
Setup: a background script screenshots every ~5 seconds to a daily folder; a cheap/fast model periodically summarizes new screenshots (~20 sampled frames per pass) into a running observations ledger; a scheduled or on-demand call to the expensive model (Fable) reviews the ledger and proposes concrete fixes.
Cost: the constant logging/summarization pass is designed to be near-free (plain text metadata plus a cheap model); the only pricier call is the periodic Fable review, run at a cadence the user controls (daily/weekly) specifically to conserve Fable session-limit budget.
Integration notes: requires OS accessibility permissions to screenshot continuously (Mac and Windows have different setup steps, per the creator, not detailed in the transcript). The creator distributed a starter script plus diagrams via the video description — not independently verified or reviewed here, since this ingest did not fetch linked video-description assets.
Try It
- Before building anything, do the cheap version manually for a day: at the end of the day, describe your last few hours to Claude/Fable from memory and ask “where did I waste time on repeat, low-value actions?” — this tests the underlying question (are there 3-4x-repeated frictions worth fixing) before investing in a screenshot pipeline.
- If building the full loop, treat it explicitly as a cost-tiering exercise: cheap/fast model for the constant summarization, expensive model only for the periodic “how do I economize” review call — the same split documented in Cost & Intelligence Levers for Agent Workflows.
- Start the observations ledger with a low bar for “confirmed pattern” (3-4 repeats) so early runs surface something actionable quickly, then raise the bar once the obvious wins are fixed.
- Treat any single-practitioner efficiency-gain claim (30 minutes/day, trending toward 2-3 hours/day) as an upper bound from an already-optimized user, not a typical expectation.
Open Questions
- The actual screenshot-capture script and the OS-specific accessibility-permission steps (Mac vs. Windows) are not detailed in the transcript — only described as available via the video description, which this ingest did not fetch.
- No detail is given on how the system avoids capturing sensitive on-screen information (passwords, private messages) before logging screenshots to disk — worth flagging before adopting this for any work touching client-confidential material.
- The “8,640 lines of metadata” and “20 sampled frames” figures are stated once, with no elaboration on how the sampling cadence was chosen or tuned.
Related
- Cost & Intelligence Levers for Agent Workflows — the cheap-executor/expensive-reviewer pattern this loop is a personal-productivity worked example of.
- Hermes + Computer Use — the acting counterpart to this passive-observation loop; contrast screenshot-as-input-to-action vs. screenshot-as-input-to-report.
- Agent Loops — the broader pattern of an interval-driven loop that reads output and reprompts; this is a personal-productivity instance of that shape.
- Prompting Claude Fable 5 — sibling practical-technique article for getting value from Fable specifically.
- Computer Use (Desktop + CLI) — the underlying screenshot/permission mechanics a capture script like this would need to work within.