Distill Habits

distill-habits — automatic habit distillation

Goal: extract the stable habits / preferences / feedback a user shows across sessions, turn them into a durable asset, and make every future conversation understand them better. Do not persist one-off task details — keep only reusable patterns.

This skill is a reflection job over your own local Claude Code logs. It reads nothing but local files and makes no network requests.

Configuration

Read config from the first file that exists, else use built-in defaults:

1. ~/.config/claude-habit-distiller/config.toml
2. /config.toml (next to this skill's install)

language        = "en"    # language for distilled habits and the summary
window_days     = 7       # look-back window when .last-run is absent
min_occurrences = 2       # how many times a signal must appear to count as a habit
max_habits      = 16      # max lines in the CLAUDE.md habits block
write_global_claude_md = true   # maintain the global habits block
exclude_projects = []     # project directory names to skip
memory_dir      = "auto"  # "auto" = detect; else an absolute path

If no config file exists, use the defaults shown above.

Path conventions

Detect the Claude home as ${CLAUDE_CONFIG_DIR:-$HOME/.claude}. All paths below are relative to it unless absolute.

• Conversation transcripts: projects//*.jsonl (one file per session, across all projects)
• User input history: history.jsonl (each line has display / timestamp (ms) / project / sessionId)
• Long-term memory dir: resolved from memory_dir. When "auto", use

projects//memory/ where ` is the entry matching the user's home path (the encoded $HOME, e.g. -Users-alice). One .md per habit, plus a MEMORY.md` index.

• Global habits block: the sentinel-managed region inside CLAUDE.md (loaded into every session)
• Last-run timestamp: /.last-run (Unix seconds; if absent, assume window_days ago)

Steps

1. Determine the time window

Read .last-run (Unix seconds). If absent, window start = window_days ago. Window end = now.

2. Collect conversation material for the window

• Find all *.jsonl transcripts under every project dir modified after the

window start (use ls -t or mtime). Skip any project in exclude_projects.

• Read them: focus on user messages, the user's corrections / feedback to

the AI, preferences the user repeats, and approaches the AI was praised for.

• Also scan history.jsonl display fields in-window: what the user says often,

which language they use, what they care about.

• With lots of material, prioritize the most recent and strongest signals — no

need to read every file word for word.

3. Distill (this is the core judgment)

Only distill stable, reusable things, in four categories matching the memory type:

• user: who the user is — language, tech-stack preferences, role, communication style.
• feedback: how the user wants the AI to work — corrections made, good approaches confirmed. Must include Why + How to apply.
• project: cross-session goals / constraints. Convert relative dates to absolute.
• reference: pointers to external resources (URLs, dashboards, docs).

Threshold: a signal must appear at least min_occurrences times, or the user must explicitly say "always do this / remember this", to count as a habit. One-off items, pure task details, and things already recorded in code/git are skipped.

4. Write to long-term memory (with provenance)

For each distilled habit, in the memory dir:

• Dedup first: if an existing file already covers it, update that file

(accumulate evidence, raise confidence) — don't create a duplicate. Delete old memories that new evidence overturns.

• File format:

```markdown --- name: description: metadata: type: user | feedback | project | reference ---

```

• Add/update a pointer line in MEMORY.md: - [Title](file.md) — hook.

5. Maintain the global CLAUDE.md habits block

Only if write_global_claude_md = true. Write the currently most useful habits (about max_habits lines, concise imperatives) into CLAUDE.md between the sentinels below. Only rewrite the content between the two markers; leave the rest of the file untouched. If the file doesn't exist, create it with just this block.

## My Habits (auto-distilled · last updated )

- 
- ...

Order by impact on day-to-day collaboration, high to low. Use language from config. Make each line actionable ("Default to Python for scripts" beats "likes Python").

6. Wrap up

• Write the current Unix seconds into .last-run (date +%s).
• Report briefly (in language): how many sessions were scanned, which habits

were added / updated / deleted, and how many lines the CLAUDE.md block now has.

Principles

• Prefer fewer, higher-confidence entries: unsure whether it's a stable habit? Skip it; wait for more evidence next time.
• Traceable: CLAUDE.md is the concise view; memory is the detailed ledger with the Why.
• Idempotent: re-running over the same window must not produce duplicates.
• Private: read only the user's own local records; send nothing over the network.

Source & license

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

• Author: betayyang
• Source: betayyang/claude-habit-distiller
• License: MIT

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