What Decluttering Prompts Weekly Actually Is
It's a curated list of prompts—single sentences or short blocks—that you feed into an LLM to reset your own workflow or study routine. The weekly cadence is mostly a marketing frame; the prompts themselves don't change much from week to week because they solve the same problems: too many half-finished files, vague todo lists, scattered notes, the general digital rot that accumulates when you spend more time in your applications than thinking about what to do next. The real value isn't the novelty of a fresh prompt. It's that someone else has already done the friction of translating a messy situation into a structured instruction a model can execute cleanly.
Getting Your Copy of Decluttering Prompts Weekly
You can grab the current batch from the site's download page. The file is typically a plain text or markdown export, roughly 15 to 40 prompts depending on which release you hit. Copy the one that matches your problem, paste it into your model of choice, adjust the input section to reflect your actual filenames or notes, and run it. No setup required beyond that. Each prompt follows the same skeleton. It asks the model to ingest a dump of your current artifacts—notes, markdown files, a directory listing, even exported chat logs—then reorganize, deduplicate, and return a clean mapping or a set of suggested moves. The most common versions I use target three situations: The output is usually a JSON or CSV you can feed back into your tool of choice. I typically import it straight into Obsidian via QuickAdd or into a Python script that performs the actual file moves.
Last November I ran the notes decluttering prompt against a vault that had been accumulating since 2019. The problem wasn't the duplicates—that part worked fine. The real issue was that roughly 40% of my files used inconsistent frontmatter keys: date_created, created, date, sometimes none at all. The prompt assumed a uniform schema and started merging files that were actually distinct drafts of the same project, simply because the timestamps overlapped within a day. I lost about six hours reconciling the merge artifacts before I realized what happened. The fix was straightforward and took me about twenty minutes: I wrote a small preprocessing step that normalized every frontmatter key into a single canonical field, collapsed empty fields, and wrote the cleaned version to a temporary folder. Only after that normalization did I run the prompt again, and the false-merge rate dropped from roughly 40% to under 3%. If you ever encounter the same issue, normalize first. Don't trust the prompt to handle inconsistent metadata. It won't.
Get the Full Details
What Beginners Miss About This Process
Most people treat the prompt as a magic reset button. It isn't. Here are two things that matter more than the prompt itself: 1. Context window waste. A lot of the prompts are written to accept large batches of text. If you throw an entire vault into one call, the model spends most of its budget summarizing noise instead of making decisions. I limit my inputs to 8,000 tokens per run and batch my directories. It sounds slower, but it actually produces cleaner output because the model isn't guessing at the structure of a dataset that's too large to reason about coherently. 2. The merge assumption. The prompts often default to merging similar items. In knowledge-work contexts, similarity doesn't mean identity. Two notes about the same topic written six months apart are not duplicates. They're iterations. Running a merge on them erases the version history that was actually useful. I override this by adding a rule to the prompt input: keep separate entries when timestamps differ by more than fourteen days and the content overlap is below 70%.
When It Doesn't Work
Decluttering Prompts Weekly breaks down in a few specific scenarios, and it's worth knowing them upfront so you don't waste time chasing results: If you hit any of these, the prompts aren't the right tool. A simple shell script with fd and bat, or a dedicated deduplication tool like dupeGuru, will handle the job faster and without hallucinated merges. Here's the exact sequence I use, from raw folder to clean output, in about twenty minutes for a typical personal vault:
First, I back up the target folder. Not a full snapshot—just a copy of the directory structure and file sizes to a date-stamped folder. If the prompt screws something up, I can restore from that in seconds without re-exporting everything. Next, I run the normalization script if my frontmatter is inconsistent. That step takes about three minutes for a 300-file vault. After that, I grab the most recent Decluttering Prompts Weekly release, copy the "notes dedup and reindex" prompt, and modify the input block to include the path to my normalized folder. I cap the model output at 4,096 tokens. Larger outputs tend to drift into repetition or start inventing file relationships that don't exist. Then I send the prompt. The model typically returns a JSON object mapping original paths to suggested destinations, along with a confidence score for each suggestion.
From there, I review the low-confidence items manually—usually fewer than ten per run—and approve the rest. I then run a Python script that executes the moves, creates a log file, and validates that the total file count hasn't changed. If the counts match, I delete the backup copy after forty-eight hours. If they don't, I revert from the backup. The whole process takes less than thirty minutes for a medium-sized vault. That's significantly faster than doing it by hand, and it catches duplicate patterns that would take me hours to spot visually.
What I Wish I'd Known Before the First Run
The biggest mistake people make is treating the prompt as a one-shot solution. It's not. Your first run will produce errors. Your second run will be better. Your third run is where the output actually becomes useful. I always keep the last three prompt versions and their corresponding outputs side by side. That way I can see what changed between runs and adjust my input format accordingly. Another thing nobody mentions: the prompts work best when you constrain the output schema yourself. The default output formats are deliberately broad. If you specify the exact keys you want back—file path, suggested action, confidence, duplicate pair IDs—the model produces cleaner, more machine-readable results. You save time downstream because the output is already in the shape you need. Decluttering Prompts Weekly is a solid starting point, but it's not magic. It's a template that requires your specific constraints, your specific data shape, and a little preprocessing before it gives reliable results. Treat it like a scaffold, not a finished product, and it'll save you real time. Ignore those two caveats about normalization and context limits, and you'll spend more time fixing the output than you would have just cleaning the files yourself.
