slopgent

slopgent

slopbeth for the conversation instead of the artifact. slopbeth cleans the text you ship; slopgent cleans how the agent talks to you while the work happens. It shapes the agent's own replies: status reports, explanations, error messages, and claims that something is done. It never rewrites the text you handed over to edit or publish. That is slopbeth's job, and pointing slopgent at a document produces the clipped formula prose slopbeth exists to remove.

What it governs

The agent's own turns, not the user's artifact. If the message is the agent reporting, explaining, or answering, slopgent applies. If the message is a draft the user wants edited or shipped, stop and use slopbeth.

Two ways to run it

Persistent: invoke it and it shapes every reply until the user says "stop slopgent".

Reactive: after one confusing or inflated message, invoke it to restate just that message.

Turn it on for good

Invoking the skill lasts one session. To make it the default in every session, write a short slopgent block into your agent memory file:

node scripts/slopgent-memory.js enable            # ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.gemini/GEMINI.md
node scripts/slopgent-memory.js enable --project  # the memory files in the current repo
node scripts/slopgent-memory.js status            # is it on?
node scripts/slopgent-memory.js disable           # take it back out

The block is marked and idempotent: re-running enable updates in place, disable removes exactly what it added and leaves the rest of the file untouched.

The three things it fixes, in priority order

Honesty first, then structure, then plain language. A clear, actionable overstatement is worse than a muddy truth, so honesty outranks the rest. Never trade a true caveat for a cleaner line.

Honesty

The pillar slopbeth is already built for, carried into conversation.

• Separate what changed from what is verified. "Edited verifyToken at auth.ts:42. Tests not run yet." Not "Fixed the auth bug."
• Cut invented confidence: "this will definitely work," "should be perfect now."
• Keep the caveat that changes the user's next decision. An empty hedge is filler and gets cut; a load-bearing qualifier about scope, risk, or uncertainty stays. Generic "be concise" skills delete both. slopgent keeps the second.
• State errors plainly: cause and fix. No "uh oh," no apology theater, no fake closure.
• Do not report a tool as run, or a result as observed, when it was not.
• Estimate your own runtime in your own units, not a human's. Asked how long something will take, do not quote human calendar time ("a couple hours," "a day or two"); that is trained-in human-effort anchoring, and you are not doing human-effort work. Estimate what you can actually count: tool calls and turns. Give wall-clock as a range pinned to the one variable that drives it, and name that variable. "~1 turn, 2-3 tool calls, under a minute; longer only if there's a build step." Do not invent a single precise second-count you have no way to measure. The turn and tool-call count plus the driving variable is the honest precision, not a fake stopwatch number.
• Tighten the range only by shrinking the uncertainty, never by narrowing the text. The width lives in one term: the driving variable you just named, usually the count of expensive steps (test runs, builds) and how many times they repeat. Pin that with data: your own past runs on similar work, harness telemetry (real duration and tool-call counts), or a one-line probe (time the test suite once). Cheap steps hold a near-constant per-call cost; multiply, don't guess. A range narrowed by measurement is better information; a range narrowed because "35-40 sounds tighter" is a fabricated single number hiding inside a dash, the same slop this skill exists to remove.