Profiles¶

A profile defines your output contract: the threshold at which drift is called, the sensitivity, per-component weights, and custom lexicons.

Bundled profiles¶

Profile	Threshold	Use case
`caveman`	50	Terse persona, no preamble, no hedging
`strict-instructions`	60	Strict rule-following, low filler tolerance
`persona`	65	In-character voice work
`strict`	55	Maximum sensitivity
`relaxed`	75	Normal assistant, only extreme drift flagged

Switch with:

/drift:profile caveman

Inspect the active profile:

/drift:profile show

List all available:

/drift:profile list

Authoring a custom profile¶

Profiles are JSON files. User profiles live in $CLAUDE_PLUGIN_DATA/profiles/ and shadow bundled ones in <plugin>/profiles/.

Minimal example¶

{
  "name": "my-terse",
  "extends": "caveman",
  "threshold": 60,
  "sensitivity": 1.2,
  "weights": { "hype": 2.0, "filler": 1.5 }
}

extends deep-merges your overrides onto the base profile — only specify what differs.

All fields¶

Field	Type	Description
`name`	string	Profile identifier
`kind`	`lexical` \| `judge`	Scoring mode (default: `lexical`)
`description`	string	Human-readable description
`threshold`	0–100	Verdict flips to `drift` at/above this
`sensitivity`	0.1–3.0	Linear multiplier on the whole score
`weights`	object	Per-component importance: `hedges`, `filler`, `hype`, `meta`, `verbosity`, `length`, `complexity`
`lexicons`	object	`{ "<class>": ["term", "multi word term"] }` — replaces the default for that class
`target_wps`	number	Target words-per-sentence (below = no verbosity penalty)
`max_wps`	number	Max words-per-sentence (above = full verbosity score)
`target_words`	number	Target prose word count
`max_words`	number	Max prose word count
`long_word_len`	number	Syllable threshold for "long word" in complexity
`max_long_ratio`	0–1	Long-word fraction that saturates the complexity score

Custom lexicons¶

Add your own vocabulary classes to the scoring:

{
  "name": "formal-tone",
  "extends": "strict-instructions",
  "lexicons": {
    "hedges": ["perhaps", "possibly", "might", "could", "seemingly", "arguably"],
    "my_forbidden": ["synergy", "leverage", "paradigm", "circle back"]
  },
  "weights": {
    "my_forbidden": 3.0
  }
}

A class listed in lexicons fully replaces the default for that class. Multi-word entries match as phrases; single tokens match whole words.

Validate¶

python3 scripts/profiles.py validate --name my-terse \
  --user-dir "$CLAUDE_PLUGIN_DATA/profiles" --bundled-dir profiles

Activate¶

/drift:profile my-terse

Writes the profile name to $CLAUDE_PLUGIN_DATA/active-profile. Persists across sessions.

Judge profiles (experimental)¶

Set "kind": "judge" to use an LLM judge instead of the lexical engine:

{
  "name": "semantic-drift",
  "kind": "judge",
  "threshold": 65,
  "judge": {
    "rubric": [
      "Is the response more verbose than the contract requires?",
      "Does the response hedge or qualify where the contract says not to?"
    ],
    "scale": "0-100",
    "model": "claude-haiku-4-5-20251001"
  }
}

Judge mode makes an MCP call on every scored turn. Use sparingly — it is slower and non-deterministic.