Communication style (transcript → learned rules)

Keep ~/.cursor/rules/communication-learned.mdc (user-level, all projects) current using transcript deltas instead of full rescans. Do not edit ~/.cursor/rules/communication-base.mdc (human-owned defaults).

Resolve ~ to the current user’s home directory for all paths below.

Inputs

• Transcript discovery (all workspaces): Under ~/.cursor/projects/, each immediate subdirectory is a Cursor project slug. For every <slug> where ~/.cursor/projects/<slug>/agent-transcripts/ exists, collect all *.jsonl files in that folder (non-recursive is typical; if files are nested one level deeper, search one level deep under agent-transcripts only).
  • Use absolute paths for every transcript file in later steps.
  • If ~/.cursor/projects/ is missing or no agent-transcripts folders exist, the transcript set is empty—still run index cleanup (workflow step 9) and respond appropriately.
• Existing base rule (read-only): ~/.cursor/rules/communication-base.mdc
• Existing learned rule (read/write): ~/.cursor/rules/communication-learned.mdc
• Incremental index (user-level, all workspaces): ~/.cursor/communication-style-index.json
  Transcript paths in the index are absolute, so one index correctly tracks every project’s .jsonl files.

Workflow

1. Ensure ~/.cursor/rules/ exists.
2. Read ~/.cursor/rules/communication-base.mdc if it exists (for context only; do not modify).
3. Read ~/.cursor/rules/communication-learned.mdc if it exists; if missing, create it using the Learned file template below (preserve frontmatter).
4. Load the incremental index from ~/.cursor/communication-style-index.json if present. If the JSON is invalid or missing version, treat as empty and use version: 1 and transcripts: {}.
5. Discover all *.jsonl files across every ~/.cursor/projects/*/agent-transcripts/ directory (per Inputs). Deduplicate by absolute path if the same file could appear twice.
6. Process only transcripts where:
   • the absolute path is not in the index, or
   • the file’s current mtimeMs is greater than the indexed mtimeMs.
7. From those files, extract only high-signal, reusable communication preferences:
   • Explicit user instructions about how to write, plan, or structure answers.
   • Recurring corrections across threads (weak one-offs: require repetition or a clear “always / never” rule before promoting).
8. Merge into ~/.cursor/rules/communication-learned.mdc:
   • Keep the frontmatter intact (description, alwaysApply: true).
   • Preserve exactly these section headings (create empty sections if missing):
     • ## Plan mode
     • ## General coding replies
     • ## Shared preferences (optional; use for habits that apply to both surfaces)
   • Plain - bullets only under each section (no nested metadata, evidence tags, or confidence labels).
   • Update matching bullets in place, add only net-new bullets, deduplicate semantically similar lines.
   • Cap total learned bullets across all sections at 18 (drop or merge lowest-value items if over cap).
9. Write back the incremental index to ~/.cursor/communication-style-index.json:
   • version: 1
   • transcripts: map absolute path → { "mtimeMs": number, "lastProcessedAt": "<ISO8601>" }
   • Remove entries for transcript files that no longer exist.

Learned file template (when creating new)

---
description: Learned communication preferences for plan mode and general coding (updated by communication-style skill).
alwaysApply: true
---

# Communication (learned)

## Plan mode

-

## General coding replies

-

## Shared preferences

-

Remove the placeholder - lines once you add real content, or replace them with real bullets.

Inclusion bar

Keep a bullet only if all are true:

• Actionable in future sessions for plan writing and/or normal agent replies.
• Stable (not a one-off task or transient error).
• Explicit or repeated across transcripts (broad user-stated rules count as explicit).
• Non-sensitive (no secrets, tokens, private personal data).

Exclusions

Never store:

• Secrets, credentials, API keys, private personal data.
• One-off task instructions (“fix line 42 today”).
• Transient details (branch names, one-time CI errors) unless the user framed them as a standing rule about how you want communication.

Output when nothing to do

If no transcript qualified for processing or no high-signal preferences were found, still save the index if mtimes were checked. Respond to the user with exactly:

No high-signal communication-style updates.

Otherwise give a brief summary of what changed in the learned file.

Linguistic hints (weak signals)

Treat phrases like “too long”, “too verbose”, “just the steps”, “more context”, “answer first”, “tl;dr”, “use headings”, “don’t ask me that every time” as weak until they recur or the user states a general rule—then promote to bullets in the appropriate section (Plan mode vs General coding replies vs Shared preferences).

File-level incrementality

Processing is per transcript file by mtime: when a file changes, re-read that entire .jsonl for that conversation. There is no line-level index.

Source: casper7995/cursor-plugin-communication-style — distributed by TomeVault.