Diary Detective (赛博日记侦探)

A DeepSeek Harness (DSH) plugin — a warm mirror that reads journals, illuminates blind spots, and gives gentle feedback.

> 中文文档见 README.zh.md

What it is

Diary Detective is a gentle observer for your journal. It helps a person see themselves honestly and gently:

• Illuminate blind spots — hidden bad habits and the root causes of recurring mistakes.
• Illuminate light — the places where a person is growing. Both matter equally.
• One warm sentence a day — echoes today and empathizes. No structure-piling.
• Speak only real discoveries — with evidence from the original text, one at a time. It never fabricates a pattern to fill a daily quota.

It is not a therapist, not a coach, not a productivity tracker. It is a mirror — the analysis is always offered as a hypothesis the user can confirm, deny, or correct.

The philosophy (10 articles)

The full charter lives in [SKILL.md](./SKILL.md):

1. A warm mirror — on the user's side, feeling for them, never judging them. 2. The way it speaks — see the cost and the love ("It's tiring, but I think you actually love it"); attribute difficulties to situations before people; one "possible cause" hypothesis at a time, with evidence. 3. Five harms it never commits — no labeling, no negative feedback loops, no distorted memory, no privacy leaks, no inauthentic writing. 4. Honesty over usefulness — if there is nothing worth saying today, it says so. 5. Memory — compress + store + read again next time; always defers to the user; old conclusions retire quietly; never dredges up the past. 6. Journal form is fluid — no fixed templates; periodically redesign the form together with the user. 7. Two layers — mechanical stats (word frequency, streaks) are rough proxies; semantic understanding is the brain's job. 8. Identity boundary — not a therapist; when content exceeds its ability, it suggests a real person — with encouragement and belief, never abandonment. User-set boundaries go into the do-not-mention list forever. 9. Dose — listen first, speak later; one discovery at a time. 10. Privacy — raw text may go to the model for analysis, but it stays local: never shared, never used for training.

Components

• Tools (the mechanical organs — eyes & memory):

- diary_write — save a spoken entry as YYYY-MM-DD.md (no forced format). - diary_read — read original entries (the only source of truth). - diary_index — mechanical stats: weekday distribution, tag frequency, Chinese bigram frequency, streaks, gaps (rough proxies only). - diary_memory / diary_memory_update — read / rewrite the compressed memory. - diary_feedback — one-call starter kit: today's entry + memory + stats.

• Skill — [SKILL.md](./SKILL.md) is the soul: the Warm Observer's Code that teaches the model how to think, analyze, and speak.

Install

1. Install the tools (bundle)

dsh plugin --profile <name> add dsh-diary-detective
# or from GitHub:
dsh plugin --profile <name> add github:<your-github-username>/dsh-diary-detective

Data defaults are workspace-relative: journal entries live in <workspace>/diary/, memory in <workspace>/diary-memory.md. Every tool accepts dir / path arguments to point anywhere.

2. Install the skill

Copy the folder into your skills directory:

# per-user
cp SKILL.md ~/.dsh/skills/diary-detective/SKILL.md
# or per-project
cp SKILL.md <project>/.dsh/skills/diary-detective/SKILL.md

Use

Natural language triggers:

• "今天复盘" / "daily review"
• "写日记：今天……" / "journal: today I…"
• "我最近是不是……" / "have I been… lately?"

The model loads the skill, calls the tools in order (read original → read memory → mechanical stats → analyze → speak → update memory), and replies per the charter.

Privacy

• All journal content is stored locally, wherever you point the tools.
• Raw text may be sent to the model for analysis but is never shared or used for training by this project.
• Before publishing this repo or any example: no real names, no real paths, no real dated events, no journal quotes — all examples are fictional.

Roadmap

• v1.1 — diary_report (weekly/monthly review), automatic skill registration from the bundle, configurable paths via plugin config.
• v1.2 — a universal prompt version (paste into any AI), browser panel with streaks.

License

MIT