Fmsi

FMSI — Find My Similar Ink

Before starting any new task, check whether the user has already done similar work. Return a verdict so the user can decide: reuse, reference, or start fresh.

When to activate

Activate automatically when the user describes a new task, project, or request — before doing the work. Also activate on explicit /FMSI, "я делал это раньше?", "проверь делал ли я такое", or any question about prior work.

Do not activate for: trivial questions, pure conversation, or clarifications about the current task that don't introduce new scope.

The four sources

Search all four, in this order. Stop early if you find an exact match.

1. Past agent chats — ~/.cursor/projects/*/agent-transcripts/*.jsonl
2. Projects on disk — ~/Desktop, ~/Documents, ~/Projects
3. Git history — commit messages and branch names in known repos
4. Existing skills — ~/.cursor/skills/*/SKILL.md and the cursor-skills repo

Workflow

1. Extract the task signature

From the user's request, pull:

• Verbs — what action (refactor, build, generate, deploy, scrape…)
• Nouns/domains — what subject (users, payments, videos, docs…)
• Tech — languages, frameworks, tools mentioned
• Artifact — what's produced (script, skill, README, API, video…)

Keep the signature to ~5–10 keywords. Discard filler.

2. Search past agent chats

# Find chat files modified in the last 90 days that mention any keyword.
rg -l --max-filesize 2M -i "<keyword1>|<keyword2>" \
  ~/.cursor/projects/*/agent-transcripts/*.jsonl 2>/dev/null | head -20

For each hit, read enough to judge relevance — usually the first user message of the chat and the last assistant message. Skip files over 2 MB (likely binary dumps).

3. Search projects on disk

# Project folders (heuristic: contains .git, package.json, pyproject.toml,
# Cargo.toml, or *.xcodeproj).
ls -d ~/Desktop/*/ ~/Documents/*/ ~/Projects/*/ 2>/dev/null | head -50

# Match by folder name first.
ls -d ~/*/<*keyword*> 2>/dev/null

# Then peek at README / package metadata in candidate folders.
rg -l --max-depth 2 -i "<keyword>" ~/Desktop ~/Documents 2>/dev/null | head -20

4. Search git history

For each known repo (start with the workspace, then ~/Desktop/*, ~/Documents/* if they have .git):

git log --all --oneline --since="12 months ago" \
  -E --grep="<keyword1>|<keyword2>" -i | head -20
git branch -a | rg -i "<keyword>"

5. Search existing skills

rg -l -i "<keyword>" ~/.cursor/skills/*/SKILL.md \
  "/Users/magic/Desktop/My Skils"/*/SKILL.md 2>/dev/null

6. Score each find

For every hit, assign:

• Match type — exact (same task + same domain), similar (overlap in 2+ of: verb, domain, tech, artifact), or tangential (1 overlap).
• Confidence — high / medium / low.
• Where — path or chat ID.
• When — date if available.

7. Return the verdict

Use this exact format (in Russian unless the user wrote in English):

**Вердикт: <делал | делал похоже | не делал>**

Уверенность: <высокая | средняя | низкая>

<Если "делал" или "делал похоже":>
Найдено:
- <одна строка на каждое релевантное совпадение: дата — где — кратко что это>

<Если "делал похоже": одна строка о том, чем похожая работа отличается от текущей.>

<Если "не делал": одна строка с предложением — начать с нуля, или попросить пример.>

Keep the verdict under 8 lines. The user can ask for details.

Scoring rules

• делал — at least one exact match: same verb AND same domain (e.g., "refactor + payments" matches a past "refactor + payments").
• делал похоже — at least one similar match, no exact: 2+ of {verb, domain, tech, artifact} overlap.
• не делал — only tangential or no hits.

Anti-patterns

• Don't search only one source and call it done — all four matter.
• Don't dump every match — only the top 3–5 most relevant.
• Don't skip the verdict line. The user wants an answer, not a list.
• Don't read entire chat files; sample the start and end.
• Don't claim "делал" without a concrete path or chat reference.

Verification

Before returning the verdict, confirm:

• Searched all four sources, not just one.
• Every "делал" / "делал похоже" claim has a concrete reference (path, chat ID, commit hash, or skill name).
• Verdict line is present and uses one of the three exact words.
• Confidence is stated.
• Total response is under 8 lines unless the user asked for details.

Additional resources

• For detailed search strategies per source and edge cases, see reference.md.