Skip to content

Self-evolving second brain with 35 AI skills, 10 agents, and people CRM. Closed-loop harness: a V-model verification lifecycle where the worker never grades its own homework. Plus paired anti-slop design skills for marketing and product UI. Works with Claude Code, Cursor, Kiro, Gemini CLI, Codex.

feat: 80% ASD-STE100 writing target and output format ladder in no-ai-slop (v3.15.0)

Huy Tieu8be6ee98025b
diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index f7bac41..d9b6d12 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
{
"name": "cog-second-brain",
- "version": "3.14.0",
+ "version": "3.15.0",
"description": "Self-evolving second brain with 36 AI agent skills, 10 agents, and people CRM. Closed-loop harness release: a V-model verification lifecycle (spec to evidence, criterion-traced checkpoints), read-only verifier agents that observe artifacts instead of trusting worker summaries, risk lanes, multi-session ultragoals with a north-star acceptance gate, plus craft skills for writing, visual work, and a paired anti-slop design system for marketing and product UI. Inspired by Garry Tan's gstack specialist sessions, gbrain knowledge patterns, and dwarves-kit closed-loop execution. Works with Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, Codex, and any markdown-reading agent.",
"author": {
"name": "Huy Tieu",
diff --git a/.claude/skills/no-ai-slop/SKILL.md b/.claude/skills/no-ai-slop/SKILL.md
index e3c9349..5272ed5 100644
--- a/.claude/skills/no-ai-slop/SKILL.md
+++ b/.claude/skills/no-ai-slop/SKILL.md
@@ -1,6 +1,6 @@
---
name: no-ai-slop
-description: Edit drafts into sharper, more human writing while preserving the writer's personal voice, or detect AI-slop patterns without rewriting. Use when the user wants a draft clearer, more direct, more opinionated, or less AI-sounding, or asks whether writing reads as AI.
+description: Edit drafts into sharper, more human writing while preserving the writer's personal voice, or detect AI-slop patterns without rewriting. Use when the user wants a draft clearer, more direct, more opinionated, or less AI-sounding, or asks whether writing reads as AI. Also the default writing standard for all agent output: 80% ASD-STE100 controlled language plus the format ladder (prose, diagram, HTML page, explainer video).
---
# No AI slop
@@ -39,6 +39,34 @@ If the goal is unclear, ask what the reader should think, feel, or do after read
- **Preserve useful edge and character.** Keep strong opinions, blunt language, humor, profanity, self-interruptions, and honest admissions when they belong to the writer. Don't replace them with safer or more professional wording.
- **Keep structure unless it's hurting the piece.** Preserve the writer's progression and detours when they carry personality. If you reorganize, say why in the What changed section.
+## Controlled-language target (ASD-STE100, adopted 2026-10-02)
+
+Default target for every explanation, report, spec, chat reply, and ticket the agent writes: about 80% of the way to ASD-STE100 (Simplified Technical English, the controlled language from aerospace maintenance manuals). Source: Karpathy, 2026-10-02, https://x.com/karpathy/status/2105819303471976479. He found STE output more readable than default model prose and softens it to "80% of the way" because the full spec is stringent. Models know the spec, so the request "write this in 80% ASD-STE100" works on any agent.
+
+The rules that carry over at 80%:
+- Procedural sentence (an instruction): 20 words max, one instruction per sentence, imperative verb first. "Run the setup script after each clone."
+- Descriptive sentence: 25 words max. Paragraph: 6 sentences max, one topic.
+- Active voice. Passive only when the actor is unknown or irrelevant.
+- Verbs: present, simple past, and future. Avoid progressive ("is running" becomes "runs") and avoid `-ing` words used as verbs or nouns where a finite verb works.
+- One word, one meaning. Pick the term once and repeat it; this is the STE version of Synonym cycling below.
+- Noun clusters of three words max. "Test run report export button" becomes "the button that exports the test run report".
+- Keep articles (the, a) and connectors that a terse model drops. Telegraphic text is shorter but slower to read.
+- Safety and warning text: the command first, then the reason. "Do not delete the lock file. Git may be writing to it."
+- Approved simple words over long ones: "use" not "utilize", "start" not "commence", "before" not "prior to", "about" not "approximately", "help" not "facilitate", "to" not "in order to".
+
+The 20% we keep outside STE: technical terms, product names, and code identifiers as they are; the writer's own voice when editing a human draft (Editing principles win over STE for a personal blog, chat, or email voice); hedges that express real uncertainty. Do not apply STE word limits to quoted material or code.
+
+## Output format ladder (adopted 2026-10-02)
+
+Same source. When the job is to make the reader understand something, prose is the lowest rung. Pick the highest rung the content and the medium support:
+
+1. STE prose (above), for answers of a few sentences and anything going into Slack, email, Jira, or a commit.
+2. A diagram (ASCII in chat, drawio or inline SVG in files), when the content has structure: a flow, a sequence, an architecture, a comparison of states.
+3. An HTML page (Artifact or local `.html`), when the reader needs to explore: tables with drill-down, toggles, an interactive model, animation of a process.
+4. An explainer video (the `release-video` skill, 3b1b-style animation with ElevenLabs narration), when the topic is a mechanism a viewer should watch unfold.
+
+Discardable artifacts are fine. A one-off web page or video that exists to explain one thing is worth building now that code is cheap. Offer the next rung up in one line when it would help and the user did not ask for it; do not build a page or video unasked for a question that two sentences answer. Every rung follows the rest of this skill: diagram node names, page headings, and video captions get the same slop rules as prose.
+
## Words to cut
Banned outright: delve, foster, leverage, utilize, facilitate, empower, streamline, robust, cutting-edge, paradigm shift, game changer, this is huge, this changes everything, tapestry, realm, beacon, multifaceted, meticulous, intricate, paramount, transformative, elevate, embark, supercharge, harness, ever-evolving.
@@ -209,6 +237,7 @@ The reliable composition order is finding -> evidence -> reasoning -> decision,
1. Read the full draft before editing.
2. Identify the core point and 3-5 voice signals to preserve, such as vocabulary, cadence, bluntness, humor, uncertainty, or digressions. Keep this note internal. If you cannot identify the core point, ask the user.
3. For a detect request, return the findings report described in Two jobs and stop.
-4. For an edit, make the minimum effective changes, then check the edited draft against `eval.md` yourself.
-5. If any check fails, fix the draft and run the checks again.
-6. Output the full edited draft and a short **What changed** section.
+4. For agent-authored text (not a human draft), write to the Controlled-language target and pick the rung from the Output format ladder before writing.
+5. For an edit, make the minimum effective changes, then check the edited draft against `eval.md` yourself.
+6. If any check fails, fix the draft and run the checks again.
+7. Output the full edited draft and a short **What changed** section.
diff --git a/.claude/skills/no-ai-slop/eval.md b/.claude/skills/no-ai-slop/eval.md
index fef4902..7957595 100644
--- a/.claude/skills/no-ai-slop/eval.md
+++ b/.claude/skills/no-ai-slop/eval.md
@@ -34,6 +34,12 @@ For detect requests, make sure the response names each pattern found with a quot
9. For model output: is there at most one parenthetical per paragraph, no semicolon-chained clauses, and prose where bullets were carrying a line of reasoning?
+## Controlled language and format
+
+1. For agent-authored text: are instructions 20 words max with one instruction each, descriptive sentences 25 words max, paragraphs 6 sentences max, voice active, and each term used with one meaning?
+2. Are noun clusters three words max, with articles and connectors kept?
+3. Is the output on the right rung of the format ladder, or does the reply offer the next rung in one line when a diagram, page, or video would explain it better?
+
## Structural slop
1. Are invented frameworks, gates, layers, pillars, and numbered taxonomies removed unless they exist in the source or genuinely reduce complexity?
diff --git a/.cursorrules b/.cursorrules
index 49fea9f..793ac63 100644
--- a/.cursorrules
+++ b/.cursorrules
@@ -55,4 +55,4 @@ For factual statements in durable notes, include:
## Response Style (ALWAYS APPLY)
-Optimize for information gain, not apparent completeness. Start with the answer or strongest finding. Never invent named frameworks, gates, layers, pillars, or numbered taxonomies unless they exist in the source material. Headings name subject matter, never rhetorical function (banned: "Why this matters", "The key insight", "What this is not", "The bottom line"). No straw-man contrasts ("It's not X, it's Y") unless X is a position someone actually holds. Space proportional to importance; every paragraph must add evidence, mechanism, example, implication, or decision. Compose as finding -> evidence -> reasoning -> decision. Stop when useful information is exhausted. Full pattern list: `.claude/skills/no-ai-slop/SKILL.md` § Structural slop.
+Optimize for information gain, not apparent completeness. Start with the answer or strongest finding. Write at 80% ASD-STE100: instructions 20 words max, descriptive sentences 25 words max, paragraphs 6 sentences max, active voice, one term per meaning. When the goal is understanding, prefer a diagram over prose, an HTML page over a diagram, and an explainer video over a page, as the medium allows. Never invent named frameworks, gates, layers, pillars, or numbered taxonomies unless they exist in the source material. Headings name subject matter, never rhetorical function (banned: "Why this matters", "The key insight", "What this is not", "The bottom line"). No straw-man contrasts ("It's not X, it's Y") unless X is a position someone actually holds. Space proportional to importance; every paragraph must add evidence, mechanism, example, implication, or decision. Compose as finding -> evidence -> reasoning -> decision. Stop when useful information is exhausted. Full pattern list: `.claude/skills/no-ai-slop/SKILL.md` § Structural slop.
diff --git a/AGENTS.md b/AGENTS.md
index 0962a2c..b3cb512 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -643,6 +643,7 @@ The following 7 skills raise output quality on writing and visual work. They enc
2. **Detect mode:** flags slop patterns without rewriting
3. Kills fake-profound kickers, summary-recap endings, hedge stacking, and formatting theater
4. Enforces concreteness, named sources, and active voice
+5. Sets the default for agent-authored text: 80% ASD-STE100 controlled language and a format ladder (prose, diagram, HTML page, explainer video), from [Karpathy, 2026-10-02](https://x.com/karpathy/status/2105819303471976479)
**Attribution:** Vendored from [petergyang/no-ai-slop](https://github.com/petergyang/no-ai-slop) (MIT). `LICENSE` and `SOURCE.md` ship alongside it.
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 126dc42..f61d91b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,18 @@
All notable changes to COG (Cognition + Obsidian + Git) will be documented in this file.
+## [3.15.0] - 2026-10-02
+
+### Added
+
+#### Controlled-language writing target and output format ladder
+Adapted from [Andrej Karpathy's post of 2026-10-02](https://x.com/karpathy/status/2105819303471976479) on how to read model output faster.
+
+- `no-ai-slop` gains a **Controlled-language target**: agent-authored text aims for about 80% of ASD-STE100, the controlled English from aerospace maintenance manuals. Instructions are 20 words max with one action each, descriptive sentences 25 words max, paragraphs 6 sentences max, active voice, one term per meaning, noun clusters three words max, articles kept. Technical terms, code, quotes, and the writer's own voice in edits stay outside the limits.
+- `no-ai-slop` gains an **Output format ladder**: when the goal is understanding, the agent picks the richest rung the content and medium support (prose, diagram, HTML page, explainer video via `release-video`) and offers the next rung in one line instead of building it unasked.
+- `eval.md` gains three checks for sentence limits, term consistency, and format choice.
+- `CLAUDE.md` and `.cursorrules` Response Style carry both rules, so agents that never load the skill still follow them.
+
## [3.14.0] - 2026-09-28
### Added
diff --git a/CLAUDE.md b/CLAUDE.md
index 25b6f02..ef2001e 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -7,6 +7,8 @@
Optimize for **information gain, not apparent completeness**. The failure mode is framework slop: ordinary reasoning dressed as a consulting memo. Full pattern list: `.claude/skills/no-ai-slop/SKILL.md` § Structural slop.
- Start with the answer or strongest finding; no introduction announcing how you will answer.
+- Write at 80% ASD-STE100: instructions 20 words max with one instruction each, descriptive sentences 25 words max, paragraphs 6 sentences max, active voice, one term per meaning, noun clusters three words max, keep articles. Technical terms and the writer's own voice in edits stay as they are. Rules: `.claude/skills/no-ai-slop/SKILL.md` § Controlled-language target.
+- When the goal is understanding, climb the format ladder: STE prose, then a diagram, then an HTML page, then an explainer video. Use the highest rung the content and medium support, and offer the next rung in one line when it would help. § Output format ladder in the same skill.
- Never invent named frameworks, gates, layers, pillars, lenses, or numbered taxonomies unless they exist in the source material or the categorization materially simplifies a complex subject.
- No sections for 1-2 paragraphs; default to continuous prose with occasional descriptive headings.
- Headings identify subject matter ("Authentication"), never rhetorical function. Banned: "What this is not", "Why this matters", "The key insight", "The real opportunity", "The bottom line", "The deeper point", "The uncomfortable truth".
diff --git a/COG-VERSION b/COG-VERSION
index f982feb..f02113f 100644
--- a/COG-VERSION
+++ b/COG-VERSION
@@ -1 +1 @@
-3.14.0
+3.15.0
diff --git a/marketplace-entry.json b/marketplace-entry.json
index 99f58b3..184d97a 100644
--- a/marketplace-entry.json
+++ b/marketplace-entry.json
@@ -7,7 +7,7 @@
"url": "https://github.com/huytieu"
},
"description": "Self-evolving second brain with 36 AI skills, 10 agents, and people CRM. Closed-loop harness release: V-model verification lifecycle with criterion-traced evidence, read-only verifier agents, risk lanes, multi-session ultragoals with a north-star acceptance gate, craft skills for writing and visuals, and paired anti-slop design skills for marketing and product UI. Inspired by Garry Tan's gstack and gbrain, and dwarves-kit. Works with Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, Codex.",
- "version": "3.14.0",
+ "version": "3.15.0",
"keywords": [
"second-brain",
"garry-tan",
diff --git a/plugin.json b/plugin.json
index f5585f0..d7d6e03 100644
--- a/plugin.json
+++ b/plugin.json
@@ -1,7 +1,7 @@
{
"$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
"name": "cog-second-brain",
- "version": "3.14.0",
+ "version": "3.15.0",
"description": "Self-evolving second brain with 36 AI agent skills, 10 agents, and people CRM. Closed-loop harness: a V-model verification lifecycle (spec to evidence, criterion-traced checkpoints), read-only verifier agents that observe artifacts instead of trusting worker summaries, risk lanes, multi-session ultragoals with a north-star acceptance gate, plus craft skills for writing, visual work, and a paired anti-slop design system for marketing and product UI. Works with Claude Code, Antigravity, Cursor, Kiro, Gemini CLI, Codex, and any Agent Plugins conformant client.",
"author": {
"name": "Huy Tieu",
diff --git a/skills/no-ai-slop/SKILL.md b/skills/no-ai-slop/SKILL.md
index e3c9349..5272ed5 100644
--- a/skills/no-ai-slop/SKILL.md
+++ b/skills/no-ai-slop/SKILL.md
@@ -1,6 +1,6 @@
---
name: no-ai-slop
-description: Edit drafts into sharper, more human writing while preserving the writer's personal voice, or detect AI-slop patterns without rewriting. Use when the user wants a draft clearer, more direct, more opinionated, or less AI-sounding, or asks whether writing reads as AI.
+description: Edit drafts into sharper, more human writing while preserving the writer's personal voice, or detect AI-slop patterns without rewriting. Use when the user wants a draft clearer, more direct, more opinionated, or less AI-sounding, or asks whether writing reads as AI. Also the default writing standard for all agent output: 80% ASD-STE100 controlled language plus the format ladder (prose, diagram, HTML page, explainer video).
---
# No AI slop
@@ -39,6 +39,34 @@ If the goal is unclear, ask what the reader should think, feel, or do after read
- **Preserve useful edge and character.** Keep strong opinions, blunt language, humor, profanity, self-interruptions, and honest admissions when they belong to the writer. Don't replace them with safer or more professional wording.
- **Keep structure unless it's hurting the piece.** Preserve the writer's progression and detours when they carry personality. If you reorganize, say why in the What changed section.
+## Controlled-language target (ASD-STE100, adopted 2026-10-02)
+
+Default target for every explanation, report, spec, chat reply, and ticket the agent writes: about 80% of the way to ASD-STE100 (Simplified Technical English, the controlled language from aerospace maintenance manuals). Source: Karpathy, 2026-10-02, https://x.com/karpathy/status/2105819303471976479. He found STE output more readable than default model prose and softens it to "80% of the way" because the full spec is stringent. Models know the spec, so the request "write this in 80% ASD-STE100" works on any agent.
+
+The rules that carry over at 80%:
+- Procedural sentence (an instruction): 20 words max, one instruction per sentence, imperative verb first. "Run the setup script after each clone."
+- Descriptive sentence: 25 words max. Paragraph: 6 sentences max, one topic.
+- Active voice. Passive only when the actor is unknown or irrelevant.
+- Verbs: present, simple past, and future. Avoid progressive ("is running" becomes "runs") and avoid `-ing` words used as verbs or nouns where a finite verb works.
+- One word, one meaning. Pick the term once and repeat it; this is the STE version of Synonym cycling below.
+- Noun clusters of three words max. "Test run report export button" becomes "the button that exports the test run report".
+- Keep articles (the, a) and connectors that a terse model drops. Telegraphic text is shorter but slower to read.
+- Safety and warning text: the command first, then the reason. "Do not delete the lock file. Git may be writing to it."
+- Approved simple words over long ones: "use" not "utilize", "start" not "commence", "before" not "prior to", "about" not "approximately", "help" not "facilitate", "to" not "in order to".
+
+The 20% we keep outside STE: technical terms, product names, and code identifiers as they are; the writer's own voice when editing a human draft (Editing principles win over STE for a personal blog, chat, or email voice); hedges that express real uncertainty. Do not apply STE word limits to quoted material or code.
+
+## Output format ladder (adopted 2026-10-02)
+
+Same source. When the job is to make the reader understand something, prose is the lowest rung. Pick the highest rung the content and the medium support:
+
+1. STE prose (above), for answers of a few sentences and anything going into Slack, email, Jira, or a commit.
+2. A diagram (ASCII in chat, drawio or inline SVG in files), when the content has structure: a flow, a sequence, an architecture, a comparison of states.
+3. An HTML page (Artifact or local `.html`), when the reader needs to explore: tables with drill-down, toggles, an interactive model, animation of a process.
+4. An explainer video (the `release-video` skill, 3b1b-style animation with ElevenLabs narration), when the topic is a mechanism a viewer should watch unfold.
+
+Discardable artifacts are fine. A one-off web page or video that exists to explain one thing is worth building now that code is cheap. Offer the next rung up in one line when it would help and the user did not ask for it; do not build a page or video unasked for a question that two sentences answer. Every rung follows the rest of this skill: diagram node names, page headings, and video captions get the same slop rules as prose.
+
## Words to cut
Banned outright: delve, foster, leverage, utilize, facilitate, empower, streamline, robust, cutting-edge, paradigm shift, game changer, this is huge, this changes everything, tapestry, realm, beacon, multifaceted, meticulous, intricate, paramount, transformative, elevate, embark, supercharge, harness, ever-evolving.
@@ -209,6 +237,7 @@ The reliable composition order is finding -> evidence -> reasoning -> decision,
1. Read the full draft before editing.
2. Identify the core point and 3-5 voice signals to preserve, such as vocabulary, cadence, bluntness, humor, uncertainty, or digressions. Keep this note internal. If you cannot identify the core point, ask the user.
3. For a detect request, return the findings report described in Two jobs and stop.
-4. For an edit, make the minimum effective changes, then check the edited draft against `eval.md` yourself.
-5. If any check fails, fix the draft and run the checks again.
-6. Output the full edited draft and a short **What changed** section.
+4. For agent-authored text (not a human draft), write to the Controlled-language target and pick the rung from the Output format ladder before writing.
+5. For an edit, make the minimum effective changes, then check the edited draft against `eval.md` yourself.
+6. If any check fails, fix the draft and run the checks again.
+7. Output the full edited draft and a short **What changed** section.
diff --git a/skills/no-ai-slop/eval.md b/skills/no-ai-slop/eval.md
index fef4902..7957595 100644
--- a/skills/no-ai-slop/eval.md
+++ b/skills/no-ai-slop/eval.md
@@ -34,6 +34,12 @@ For detect requests, make sure the response names each pattern found with a quot
9. For model output: is there at most one parenthetical per paragraph, no semicolon-chained clauses, and prose where bullets were carrying a line of reasoning?
+## Controlled language and format
+
+1. For agent-authored text: are instructions 20 words max with one instruction each, descriptive sentences 25 words max, paragraphs 6 sentences max, voice active, and each term used with one meaning?
+2. Are noun clusters three words max, with articles and connectors kept?
+3. Is the output on the right rung of the format ladder, or does the reply offer the next rung in one line when a diagram, page, or video would explain it better?
+
## Structural slop
1. Are invented frameworks, gates, layers, pillars, and numbered taxonomies removed unless they exist in the source or genuinely reduce complexity?