← Back to Skills Marketplace
239
Downloads
0
Stars
0
Active Installs
7
Versions
Install in OpenClaw
/install talk-normal
Description
Stop LLM slop. A curated system prompt that cuts verbose, corporate-sounding LLM output by 56-71% (measured) while preserving information. Works bilingually...
Usage Guidance
This skill is internally consistent but makes a persistent change to your agent configuration. Before installing: 1) Inspect install.sh and prompt.md locally (they are short and readable). 2) Back up your AGENTS.md (it may be in the current directory, $HOME, or the path given by OPENCLAW_WORKSPACE). 3) Confirm you want a global, always-on prompt rule applied to agents that read AGENTS.md; uninstall with bash install.sh --uninstall if you change your mind. 4) Avoid running any external update automation unless you trust the upstream repo; manual review of updates is recommended. Running the provided script is a local file edit only—no credentials or network calls are involved.
Capability Analysis
Type: OpenClaw Skill
Name: talk-normal
Version: 0.6.2
The skill is a utility designed to reduce LLM verbosity and 'slop' by injecting stylistic instructions into the OpenClaw AGENTS.md configuration file. The install.sh script performs transparent, idempotent file modifications to manage the prompt block, and the prompt.md file contains only linguistic constraints (e.g., avoiding filler, direct phrasing) without any indicators of data exfiltration, malicious execution, or harmful prompt injection.
Capability Assessment
Purpose & Capability
The name/description (reduce LLM verbosity via a curated system prompt) matches what the bundle contains: a prompt.md and an installer that inserts that prompt into AGENTS.md. No unrelated binaries, credentials, or surprising capabilities are requested.
Instruction Scope
SKILL.md and install.sh limit actions to locating/creating/modifying AGENTS.md (searching ., $HOME, and ${OPENCLAW_WORKSPACE}). The installer appends or replaces a marked block with the contents of prompt.md. It does not read arbitrary files, call remote endpoints, or exfiltrate data.
Install Mechanism
There is no remote install step or external download. This is instruction-only with a local install.sh included; running the script performs only local file edits. No third-party package installs or network fetches are performed by the provided script.
Credentials
The skill declares no required environment variables or credentials. The installer optionally reads OPENCLAW_WORKSPACE to locate AGENTS.md but does not require secrets or keys.
Persistence & Privilege
Although the registry flag always:false, the installer intentionally injects an always-on behavior block into AGENTS.md so the change is persistent across conversations until you uninstall. This is expected for the skill's purpose but is a persistent, workspace-wide change to agent behavior and should be acknowledged before running.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install talk-normal - After installation, invoke the skill by name or use
/talk-normal - Provide required inputs per the skill's parameter spec and get structured output
Version History
v0.6.2
Add prompt-chatgpt.md, a 1400-char compressed variant for ChatGPT's custom-instructions 1500-char field limit. Responds to a real bug report from @ppdng on Twitter (prompt.md v0.6.1 at 3527 chars doesn't fit). prompt.md itself is unchanged — the ClawHub install flow for OpenClaw users is unaffected. Use prompt.md for OpenClaw / API / Cursor; use prompt-chatgpt.md only for the ChatGPT custom-instructions field. All load-bearing rules preserved in the compressed variant.
v0.6.1
Rule 3 (closing-stamp ban) goes structural. v0.6.0 Andreessen test dropped negation-frame violations to 0 but surfaced a new variant 一句话落地 as a section header — the model pivoted from the banned 一句话总结 to an unlisted synonym. v0.6.1 adds an explicit intent statement, expanded word list (一句话落地, 一句话讲, 一句话概括, 一句话说, 一句话收尾), and structural patterns (一句话X：, X一下：) so future variants are caught before they appear. Same word-list-to-structural arc rule 2 went through in v0.5.0.
v0.6.0
Add version tracking to install flow. prompt.md now has <!-- talk-normal X.Y.Z --> as first line. install.sh prints the version and absolute target path on install/update. sync-skill.sh auto-reads version from prompt.md and bumps skill/SKILL.md. Diagnostic purpose: users in multi-workspace setups can grep AGENTS.md for the version currently active and read install.sh output to confirm which source file was used. No rule-logic changes — this is an install-UX release.
v0.5.0
Response to v0.4.0 GPT-5.4 leak test (3 classic + 1 new variant + 1 closing-stamp bypass). Four changes: (1) broadened rule 2 to cover any order/variant, not just 'open by negating'; (2) removed the specific '不是交易信号' example from rule body because the model copied it verbatim in v0.4.0; (3) added four BAD/GOOD example pairs below the meta constraint for positive demonstration; (4) promoted the closing-pattern rule from position 13 to position 3 so Chinese summary stamps like 一句话总结 get comparable attention. See regressions/rule-17-negation-frame.md for full leak history (many → 6 → 4 → 3 → target 0-1).
v0.4.0
Promoted the negation-frame rule from position 14 to position 2 and added a top-of-prompt meta constraint marking it as the hardest rule. Added chained-negation variant coverage (不是A不是B而是C). Extended closing-pattern rule to also ban Chinese summary stamps (一句话总结, 总结一下, 简而言之, 概括来说, 总而言之). Motivated by observed GPT-5.4 leak test: 6 violations in v0.3.0, 4 in v0.3.1, targeting 0-1 in v0.4.0. See regressions/rule-17-negation-frame.md in the upstream repo.
v0.3.1
Stop hardcoding rule count in skill description. Rule count changes with every release; description now uses 'curated system prompt' and the rule summary is grouped into categories instead of an enumerated list.
v0.3.0
Initial ClawHub release. 17 rules, bilingual (English + Chinese), idempotent installer, measured 56-71% character reduction on GPT-5.4 / GPT-4o-mini across 10 test prompts without losing information.
Metadata
Frequently Asked Questions
What is talk-normal?
Stop LLM slop. A curated system prompt that cuts verbose, corporate-sounding LLM output by 56-71% (measured) while preserving information. Works bilingually... It is an AI Agent Skill for Claude Code / OpenClaw, with 239 downloads so far.
How do I install talk-normal?
Run "/install talk-normal" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is talk-normal free?
Yes, talk-normal is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does talk-normal support?
talk-normal is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created talk-normal?
It is built and maintained by hexiecs (@hexiecs); the current version is v0.6.2.
More Skills