Useful Or Cut

Useful or Cut

Based on Shreyas Doshi's flowchart: Is it useful? → Yes: take what's useful, ignore the rest. → No: ignore the rest.

The flowchart is a reader's filter. This skill runs it in reverse, against the draft, before the reader ever sees it. The agent plays hostile reader on its own output.

The core rule: the "no" branch is deletion, not revision. In the original, "no" loops back to ignore the rest. Here it means cut the sentence. Do not soften it, hedge it, or rewrite it into something adequate. Cut it.

Workflow

1. Find the actual claim

Before writing anything, state in one sentence what the user is claiming, and what a reader can do differently after reading it. Show this to the user and get agreement.

If the idea doesn't support a claim — if it's an observation, not an argument — say so. Do not manufacture an argument to fill the shape of a blog post.

2. Size the piece to the idea

Estimate the length the idea can carry at full density. State the number before drafting. Common outcomes: a sharp idea is 400–800 words; a genuinely developed argument is 1200–1800.

Never write to a target length. If the user asks for 1500 words and the idea supports 600, write 600 and say why. Padding to hit a number is the single largest source of slop.

3. Draft

Write it. Don't self-censor during drafting — the cutting pass handles that.

Use the user's own words, examples, and framings from the conversation wherever they exist. Their specifics are the part no generic model output can replicate; they are the density.

4. The cut pass (mandatory — never skip, never merge into step 3)

Go paragraph by paragraph. For each one, write in ten words or fewer what the reader takes away.

Delete the paragraph if:

• You can't complete that sentence
• The takeaway duplicates an earlier paragraph's
• The takeaway is "context," "setup," "transition," or "restates the above"
• It's true but the reader already believed it

Then go sentence by sentence inside the survivors and do it again.

Expect to cut 30–50% of the draft. If you cut less than 20%, you did the pass wrong — run it again, harder. The common failure is grading your own writing generously.

5. Known slop shapes — cut on sight

• Throat-clearing openers. "In today's fast-paced world," "We've all been there," any scene-setting before the claim. Start at the claim.
• The restating conclusion. A final section that summarizes what was just said. End on the last real point.
• "It's not X, it's Y." Fine once per piece, maybe. Three times is a tic.
• Rhetorical questions as transitions. "So what does this mean for you?" Just say the thing.
• Hedge stacks. "It's worth noting that it may in some cases be possible that…"
• Inflated lists. Six items where three are real and three were generated to reach six. Ship three.
• Symmetry padding. Giving a weak counterpoint equal space because the structure looked balanced.
• Adjective pairs. "clear and concise," "robust and scalable." Pick one or neither.
• Explaining the obvious to sound thorough. If your reader knows what an API is, don't define it.

The following six shapes are adapted from petergyang/no-ai-slop (MIT):

• Colon reveals. A noun phrase, a colon, then a dramatic reveal: "The best part: it learns." Rewrite as a plain sentence. Colons are for lists, labels, and quotes, not fake drama.
• Superficial -ing analysis. Trailing clauses that pretend to explain meaning: "highlighting," "underscoring," "reflecting," "showcasing." Replace with the concrete consequence or cut.
• Importance puffery. "Stands as a testament," "marks a pivotal moment," "plays a vital role," "underscores its significance." State the fact and let the reader judge whether it matters.
• Weasel attribution. "Experts agree," "studies show," "widely regarded as," "many argue." Name the source or cut the claim. No source? Ask the user — never invent one.
• Synonym cycling. Rotating "the agent / the assistant / the tool" for the same referent. If the clear word is right, repeat it.
• Fake-profound kickers. A final "deep" line that turns the point into a cute metaphor or mic-drop. Delete it — do not rewrite it into a better metaphor — and end on the clearest concrete sentence already in the draft.

6. Fidelity check

The draft must contain no claim the user didn't make or wouldn't endorse. Agents pad thin ideas by inventing supporting assertions, statistics, and examples. Flag anything you introduced:

> Added: the comparison to X, the figure of 40%, the anecdote about Y. Confirm or cut.

Invented numbers and fake anecdotes are worse than slop — they're wrong.

7. Self-check before the exit gate

Before handing the draft to your prose-quality gate (if you have one, e.g. an npm run slop-check script), answer each with pass/fail; fix and re-check on any fail. (Self-check pattern from petergyang/no-ai-slop's eval.md, MIT.)

1. Does every paragraph have a ten-words-or-fewer takeaway that isn't "context" or "recap"?
2. Are all fifteen slop shapes in §5 absent (or flagged deliberately)?
3. Is every claim, number, and example the user's own — or explicitly flagged as added?
4. Does the piece end on its last concrete point, not a summary or kicker?
5. Would the user recognize the draft as their own voice?

8. Report

End with a short cut log so the user can overrule you:

Claim: 
Drafted: 1,340 → Final: 780 words (42% cut)
Cut: opening scene-setter, section 3 (duplicated section 1),
     closing summary, 4 hedges
Added by me (confirm): the Postgres example in §2

What this skill does not do

Usefulness is not the only test. A piece can pass this filter and still be dishonest, unattributed, or unkind. Accuracy, sourcing, and credit are separate gates — this one only governs whether text earns its place on the page.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: pktikkani
• Source: pktikkani/agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.