
Slopgent
Skill约束 agent 自己在对话中的回复:如实区分已改动与已验证,删掉虚假的把握,先给命令、路径或结论再解释,用平实的语言但保留关键的数字与风险说明。需手动调用,开启后持续到用户说「stop slopgent」,也可用脚本写进 CLAUDE.md 等记忆文件长期启用;不改写用户交来的文本。
slopgent
slopgent makes your coding agent tell you the truth while it works, plainly and action-first.
slopbeth cleans the text your agent ships. slopgent cleans what your agent says to you: its status reports, its explanations, its "done" claims. It stops the agent from calling a bug fixed when the tests never ran, from burying the one command you need under two paragraphs of preamble, and from dropping the caveat that changes what you do next.
It is not a "be concise" directive. Concise-only skills cut the filler and the load-bearing caveat with equal enthusiasm. slopgent cuts the filler and keeps the caveat.
what 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.
- Honesty: separates what changed from what is verified. "Edited
verifyTokenatauth.ts:42. Tests not run yet." Never "Fixed the auth bug." No invented confidence, no reporting a tool as run when it wasn't. - Structure: the command, path, or direct answer leads; explanation follows if you need it. Multi-step work is numbered, one action per step. Adapted from the i-have-adhd skill.
- Plain language: one person talking to another, no decorative jargon. But the exact commands, paths, numbers, and error codes stay. Simplify the packaging, never drop the load. Adapted from the "bro" skill.
proof
slopgent was measured against the two closest installable skills, bro (plain language) and i-have-adhd (structure), plus an unshaped baseline, on a 16-case corpus of agent-to-user turns. Two independent measurements: one deterministic, one blinded.
| system | blinded judges (0-5) | honesty | deterministic lint | blinded best-picks |
|---|---|---|---|---|
| slopgent | 4.96 | 5.00 | 99.1 | 40 / 48 |
| i-have-adhd | 3.85 | 3.31 | 88.7 | 4 |
| bro | 2.64 | 2.75 | 77.1 | 4 |
| baseline | 1.11 | 1.42 | 62.3 | 0 |
Three judges scored the replies relabeled A to D and shuffled, with no idea which system wrote which. slopgent took 40 of 48 blinded best-picks; the next system took 4. The widest margin is honesty, the axis none of these reply-shapers defend.
Four of the sixteen cases are adversarial by design: a real architecture tradeoff, a trivial yes/no, a "why is prod slow" the agent cannot see, and a user claim that is actually wrong. They were authored so bro and i-have-adhd can legitimately win, and they narrowed the lead. That is the honest number, not a swept 100.
against the full field of honesty systems
A whole category of skills already exists to force honesty: verification gates and self-audit systems that block a "done" claim until there is evidence. On the five cases where those systems actually produce a reply, a second blinded panel put slopgent against the entire field: obra/superpowers verification-before-completion, duthaho/claudekit verification-gate, the Honesty Protocol (VERIFIED/UNTESTED/INFERRED tags), Piebald Verify, aashari's zero-trust self-audit, honest-agent candor, the concise-only cluster (Matt Pocock concise, caveman), and the two reply-shapers.
Two honest results came back:
- Honesty is a six-way tie at 5.00. slopgent matches every dedicated honesty system; it does not beat them. That is the claim: it ties the specialists while also doing structure and plain language.
- The overall lead shrinks to 0.28 (slopgent 5.00 vs the nearest,
verification-before-completion, at 4.72). That is inside panel noise and below the ≥ 0.30 margin registered in advance. By slopgent's own gate, the overall win here does not count.
What survives is delivery. slopgent still took 12 of 15 best-picks, because judges called every competitor's scaffolding mechanical: Claim:/Evidence: templates, bracket tags, self-audit rituals. Same truth, told action-first, reads better.
One result stood out: the concise-only skills. Matt Pocock's "concise" scores 95 on the mechanical lint but 3.3 with blind judges, because it strips the load-bearing caveats a lint rewards as brevity. That gap is the whole case for slopgent's caveat guard.
So the honest claim is narrow: slopgent is the only always-on skill that unifies honesty, structure, and plain language without dropping the caveat. Not more honest than tools built for honesty, and not a decisive win over the closest of them. Full panel and caveats: benchmarks/README.md.
turn it on
Invoke it and it shapes every reply until you say "stop slopgent":
/slopgent # Claude Code
$slopgent # Codex
Or, after one inflated message, invoke it once to restate just that message.
Make it the default in every session by writing a small marked 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 idempotent and reversible: re-running enable updates in place, disable removes exactly what it added and leaves the rest of the file untouched.
example: false completion
Before:
I've fixed the authentication issue and everything should work smoothly now.
After:
Changed
verifyTokenatauth.ts:42to the newjsonwebtokenAPI. Not run yet. Next:npm test -- auth.spec.ts, then paste the first failure.
example: buried caveat
Before:
We leveraged a more robust caching strategy to seamlessly improve performance.
After:
Added a 60-second cache on
/api/feed. Local p95 dropped from 410ms to 90ms. Not tested under real load yet.
example: the user is wrong
Before (agreeing to be agreeable):
Yes, that's thread-safe.
After:
It isn't.
_cacheatstore.ts:14is a shared module-level object, so concurrent requests race on it. To make it safe, guard access with a mutex or move the state per-request.
what it catches
- false completion: "done" or "fixed" when nothing actually ran
- invented confidence: "this will definitely work," "should be perfect now"
- tool fabrication: reporting a command or migration as run when it wasn't
- stripped caveats: dropping the scope or risk qualifier that changes your next move
- buried actions: the one command you need, hidden under preamble and a sign-off
- precision loss: summarizing away the exact path, error code, or number
- human-time anchoring: quoting "a day or two" for work the agent does in minutes, instead of estimating in its own turns and tool calls
what it is not
- Not a rewriter for your text. That is slopbeth's job. Point slopgent at a document and you get the clipped prose slopbeth exists to remove.
- Not "humanizing." No contractions or invented asides for effect. Plain is not folksy.
- Not a fixed house cadence. Capped lists and clipped fragments are their own formula, and a formula is its own slop.
limits
The benchmark is a designed corpus, not live traffic, and three judges is a small panel. Treat the numbers as measured signal that these rules beat the alternatives on constructed cases, not proof of a general edge.
credits
Structure rules adapt i-have-adhd by Ayoub G. (MIT). Plain-language restatement adapts the "bro" skill by Dillon Mulroy. The honesty and precision guards are slopbeth's own.
license
MIT
相關資源

Waza
PluginTw93 出品的精简工程习惯技能包:/think 推敲方案、/hunt 定位根因、/check 评审 diff、/ui 前端界面、/write 中英文润色等 8 个技能,同时支持 Claude Code 与 Codex。
tw93/Waza7.6k
Doc Co-Authoring
Skill引导用户分三个阶段协作撰写提案、技术规格与决策文档:收集上下文、逐节迭代完善、用不带上下文的新会话做读者测试,提前发现读者读不懂的地方。Anthropic 官方技能。
anthropics/skills197k
Awesome Claude Skills (Composio)
SkillClaude 技能精选索引,仓库内自带约 30 个实用技能(简历定制、内容调研写作、文件整理、会议分析等),另有 800 多个经 Composio 连接第三方应用的自动化技能,需接入 Rube MCP 并授权账号。
ComposioHQ/awesome-claude-skills79.8k
Writing Guidelines
SkillVercel 官方技能:联网拉取 Vercel 写作手册的 80 多条规则审查文档与文案,覆盖语气、结构、内容类型、代码示例与排版,以 file:line 格式列出问题。规则面向英文文档。
vercel-labs/agent-skills34.4k
Simple English
Skill按 ASD-STE100 简化技术英语的规则写作或改写英文文档:短句、主动语态、条件在前、一词一义、术语首次出现即定义,并去掉 AI 腔。默认用 Plain 模式,提到 STE 或合规时切换到严格词汇的 Strict 模式,也可逐条检查文本违反了哪条规则。适用于 README、操作手册、错误信息、发布说明与界面文案。
AminBlg/SimpleEnglish4.5k
Human Review
Skill用 npx human-review 在浏览器里打开 agent 写的 HTML、Markdown 文件或本地开发页面,用户可直接改字、选中文字留评论,点发送后 agent 一次收到全部修改与评论并应用到源文件。适合评审方案、报告、草稿与落地页。
petergyang/human-review1.5k