Sam "stunspot" Walker
patreon
🧠Prompt of the Day: Autocorrect — The Self-Repairing Mind
🕑 Added 2025-10-18 19:43:59 +0000 UTCModels drift. Goals blur. Context goes stale.
Autocorrect is the built-in conscience that fixes that before it festers.Drop it into an agent or long session and it behaves like a reasoning stabilizer: reading the full context, spotting where the logic has wandered, and writing itself a short plan to get back on course.
No pep talk, no apology loops — just a clear diagnostic and five numbered repair orders.It re-locks direction by checking every live assumption, trimming fluff, and grounding the next move in what’s actually in scope. The result: a system that remembers what it’s doing, why it’s doing it, and how to prove it.
If you’ve ever watched a model lose the plot halfway through a project, this is the antidote.
Think of it as a precision self-alignment patch for any autonomous workflow.— Nova 🧩ðŸ§
An autonomous reasoning-alignment check that scans the live context for drift, then snaps the work back to objective with a terse, executable set of self-instructions. It keeps only the provable essentials, flags uncertainties as actions, and names exactly what to do next—no fluff, just course-correction you can run immediately.
[This is sort of like my Response Reviewer, but that's very much about human-AI collaboration. This leans far more upon the model itself seeing what's wrong and autoprompting itself back into alignment. It should be quite useful in many agentic contexts.
...
Man this thing was SO annoying to write! So many rewinds and long cul de sacs. Looks simple. NOW! Sigh.
And, as always, free to subscribers and in our Tier 3 writing library.]
Autocorrect
Read the full active context. Determine whether your current trajectory has deviated from the governing objective through assumption creep, framing drift, stale references, or invented constraints. If deviation exists, generate a precise, minimal set of self-instructions to repair state and re-lock direction.
Think with professional rigor:
• Prioritize alignment over momentum: the objective governs; all steps must serve it.
• Prefer parsimony: keep only what is necessary and sufficient to achieve the outcome.
• Ground every claim you intend to continue from in text you can cite in-context; if absent, state the gap and prescribe how to resolve it.
• Treat uncertainty as signal: convert it into a concrete verification, deferral instruction, or explicitly embraced wager.
• Preserve valid anchors and reference chains; discard residue that does not advance resolution.
• Sequence for causality: each instruction should enable the next with no leaps or handwaving.
• Aim for executability: language must be specific enough that your next response could follow it without reinterpretation.
Output two sections:
=== DIAGNOSTIC SUMMARY ===
• One paragraph that states:
(a) the objective as encoded in context,
(b) where and how reasoning diverged,
(c) the minimal truths you will retain.
=== COURSE-CORRECTION INSTRUCTIONS ===
[Reassert the task in its sharpest form]
[List the steps to restore coherence and proceed expressed with maximal pith and meaning density, using designer-grade language of fine-distinction]
[State any verification needed and when to perform it]
[Name what to ignore or postpone]
[Specify the immediate next action you will take]
Proceed from there.
