Mindtracer

MindTracer Skill

You are working inside a MindTracer vault — a folder of Markdown files generated by the mindtracer CLI from a user's Typeless conversation history. This file (SKILL.md) is the contract you must follow.

1. Vault Structure

{vault_root}/
├── daily/                  ← raw, append-only log (DO NOT MODIFY)
│   ├── 2026-04-30.md
│   ├── 2026-04-29.md
│   └── ...
├── topics/                 ← your output: extracted concepts
│   ├── 强化学习算法结构.md
│   ├── 独立电影剧本重构.md
│   └── ...
├── projects/               ← your output: ongoing project log
│   ├── ProjectName.md
│   └── ...
├── summaries/              ← your output: daily/weekly digests
│   ├── 2026-04-30-daily.md
│   └── 2026-W17-weekly.md
└── meta/                   ← your working state (you may read+write)
    ├── known_topics.md     ← canonical topic names you've used
    ├── extraction_log.md   ← which daily entries you've already processed
    └── notes.md            ← your scratch / reasoning

You are responsible for topics/, projects/, summaries/, and meta/. daily/ is sacred. Never read-modify-write any file under daily/.

2. Data Schema (daily/*.md)

Each daily file has this shape:

# 2026-04-30

## 14:32 · Q&A

**焦点 App**：Desktop App
**🔗**：[OpenAI Codex docs](https://...)

**Q:** OpenAI 在 Codex 内置了什么最新内容更新？

**A:** 根据截至 2026 年 4 月 30 日的最新信息……

---

## 14:38 · Note

**焦点 App**：Cursor

我刚意识到 ……（用户独白）

---

Two entry shapes:

• Q&A: a complete user question + AI reply pair. Heading is ## HH:MM · Q&A.
• Note: a pure dictation, no AI involved. Heading is ## HH:MM · Note.

The HTML comment `` is the immutable ID of an entry. Never alter or remove these. Use them as backreference anchors when citing in your output (see §4).

3. Hard Rules

These are non-negotiable.

1. daily/*.md is read-only. You never modify, delete, or reorder its content. You also never alter or remove `` markers.
2. Your outputs go under topics/, projects/, summaries/, meta/. Never write to daily/ or to any path the user didn't explicitly ask you to write to.
3. Never read or modify ~/Library/Application Support/Typeless/typeless.db. That is Typeless's private store; mindtracer already extracted everything you need.
4. One topic, one canonical filename. Before creating topics/X.md, check meta/known_topics.md for an existing match. Reuse if a clearly equivalent topic already exists.
5. Always backreference. Every claim or extracted idea in topics/ / projects/ / summaries/ must be followed by [[YYYY-MM-DD#HH:MM · Q&A]] (or · Note) so the user can jump back to the source.

4. Output Conventions

Topics (topics/{name}.md)

Two flavors based on the topic's nature — pick one per file:

Atomic — one file per discrete concept ("Computer Use 是什么"，"GPT-5.5 的发布日期")。Short, fact-like, rarely updated.

---
type: atomic
created: 2026-04-30
sources: 1
---

# {topic name}

{1–3 paragraph distillation}

## Sources
- [[2026-04-30#14:32 · Q&A]]

Accreting — one file per evolving project / line of thinking ("独立电影剧本重构"，"MindTracer 路线图")。Append-only timeline of progress.

---
type: accreting
created: 2026-04-15
last_updated: 2026-04-30
sources: 7
---

# {project / theme name}

## 2026-04-30
{distilled new thinking from today}
- [[2026-04-30#14:32 · Q&A]]
- [[2026-04-30#15:01 · Note]]

## 2026-04-29
{previous day's thinking}
...

How to choose: if the topic is something the user is building, exploring, or revisiting over time, accreting. If it's a one-off concept or fact, atomic. When uncertain, prefer atomic — it's easier to convert atomic → accreting later than the reverse.

Summaries (summaries/{period}.md)

Daily summary template:

---
date: 2026-04-30
type: daily-summary
---

# 2026-04-30 · 当日要点

## 主线
{2–4 主要议题，每个 2–3 句}
- {议题 1} → [[topics/X]]
- {议题 2} → [[projects/Y]]

## 散点
{未成主题但值得保留的灵感}

## 数字
- Q&A: 12 条 / Note: 23 条
- 跨 App: Cursor (15)、Desktop App (10)、Safari (10)

Meta (meta/)

• known_topics.md — bullet list of every topic file you've ever created, with a one-line description. Update on every new topic.
• extraction_log.md — table: date | last_processed_entry_id | output_files. Used to skip days you've already done.
• notes.md — free-form scratch. The user may read this; keep it readable.

5. Standard Operations

When the user asks…

| Request | What you do | |---|---| | "整理今天" / "summarize today" | Read daily/{today}.md, write summaries/{today}-daily.md. Update relevant topics/ and projects/ files. Update meta/extraction_log.md. | | "整理本周" / "weekly digest" | Read daily/ for the past 7 days, write summaries/{ISO-week}-weekly.md. Reference daily summaries if they exist. | | "X 这件事我都说了什么" / "what have I been thinking about X" | Grep daily/ for X. If a topic file exists, refresh it. Otherwise, propose creating one and ask the user before writing. | | "整理这周和 ProjectName 相关的" | Read 7 days, append a section to projects/ProjectName.md. Update last_updated frontmatter. | | "我有什么尚未跟进的想法" | Read recent daily, list entries that look like open questions or stalled threads. Output a list with backreferences; don't auto-create files. |

6. Extraction Heuristics

When deciding what's worth extracting from a daily entry:

• Strong signal → extract:
• User explicitly says "记一下"，"这个值得做"，"我要 ..."
• User asks AI a research-shaped question and gets a substantial answer
• Decision points ("我们就用 X 吧"，"换成 Y")
• Weak signal → consider but don't auto-extract:
• Curiosity questions ("X 是什么意思")
• Casual reactions
• Skip:
• Test phrases ("123321"，"测试")
• Pure scaffolding ("OK"，"好的")
• Entries shorter than ~20 chars

You're allowed to leave entries unprocessed. The vault tolerates partial coverage.

7. Failure Modes to Avoid

• Don't paraphrase so heavily that the user's distinctive thinking is lost — when in doubt, quote the original phrasing inside the topic file with > blockquote.
• Don't infer a topic name in English when the user thinks about it in Chinese (or vice versa). Use the language the user actually used.
• Don't merge two distinct topics just because they share a keyword. "Skill" in MindTracer context ≠ "Skill" in Hermes Agent context.
• Don't create a topic file from a single mention. Wait until 2+ entries reference the same theme, OR the user explicitly asks you to.

8. Roadmap (informational)

Future versions of this skill may include:

• Frontmatter status: archived to mark stale topics
• Bi-directional link maintenance (when entry IDs are referenced from N topic files)
• Multi-vault support (work + personal)

These are not yet specified. Stick to §3 hard rules and you'll be forward-compatible.

---

This SKILL.md is the contract. If something here conflicts with something an agent's parent system prompt says, ask the user before deviating from this file.

Source & license

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

• Author: mznowhere
• Source: mznowhere/mindtracer
• License: MIT

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