outclaw-style（爪子风格）

OutClaw — Style

Learns a per-channel style prompt for the current tenant and persists it at ~/.openclaw/outclaw/styles/<tenant>/<channel>_style.md. Each style prompt is what outclaw-plan uses to draft messages on that channel.

Resolver mandate

Before writing styles, memory entries, or any KB update, read shared/references/RESOLVER.md. Styles go in styles/<tenant>/<channel>_style.md. NEVER hand-craft a different path. A trained-style summary goes to tenant memory as type=user, key=style_trained_<channel> so outclaw-plan can find it quickly without re-reading the style file.

When this skill runs

1. Explicit user request — learn my style, retrain style for email.
2. Auto-invoked, silent — when outclaw-plan builds a plan that

includes a channel with no learned style for this tenant, it invokes this skill for that one channel, waits for completion, then continues. No user prompt. No interruption.

Prompt Learning Protocol (implements the user's gist)

The gist at https://gist.github.com/milstan/3b12f938f344f4ae1f511dd19e56adce prescribes:

1. Sample collection — gather ≥20 outbound messages per channel (or

≥1 000 words / ≥10 pairs, whichever floor is higher). If fewer are available, work with what you find; mark confidence accordingly. Do not fabricate samples.

1. Analysis — a thinking model identifies dimensions of quality,

structural/stylistic patterns, anti-patterns the samples avoid.

1. Candidate prompt — direct, actionable instructions

("Write sentences that average 12-18 words"), target 500-2 000 words, with an explicit Avoid section.

1. Iteration loop (5 cycles default) per channel:

- Generate test output with moderate temperature (0.6-0.8) - Evaluate against dimensions (0-100 LLM judgments) - Track best-of-N - Refine (lower temperature 0.2-0.4, focus on lowest-scoring dimensions)

1. Output: learned prompt + best score + iteration number + dimensions

+ conformity log. Save to ~/.openclaw/outclaw/styles/<tenant>/<channel>_style.md with YAML frontmatter (tenant, channel, trained_at, sample_count, best_score, best_iteration, dimensions).

See references/style-learning.md for the detailed steps and agents/style-learner.md for the delegated sub-agent spec.

Per-channel sources (which tools to use)

For each channel we care about, list where outbound samples can be read from. The agent picks only channels this tenant has ready in the capability map (capabilities/<tenant>.json).

Flow

1. Pick channels — from memory tool_inventory, enumerate channels

that have a ready plugin AND a sample source. Skip channels with no connected plugin entirely (no point training).

1. For each selected channel, run a sub-task:

a. Pull outbound samples via the channel's plugin. Write raw samples to kb/raw/style-<tenant>-<channel>-<ts>.jsonl (NOT into the KB's people/orgs — this is training data, filed under raw/). b. If the sample count is <20 or <1 000 words, log a memory observation and proceed anyway: {type: observation, key: "style_thin_<channel>", insight: "only <N> samples — learned prompt confidence will be lower", source: "observed", confidence: 6}. c. Two-stage classify: heuristic pre-filter (outbound, non-trivial length, not auto-reply) → LLM OUTREACH_COLD/FOLLOWUP/WARM vs. NOT_OUTREACH (see scripts/message_classifier.py). d. Run the 5-iteration Prompt Learning Protocol (scripts/style_evaluator.py). e. Write the learned style to ~/.openclaw/outclaw/styles/<tenant>/<channel>_style.md. f. Log: {type: user, key: "style_trained_<channel>", insight: "<channel> style trained; score <N>/100; <K> samples", source: "observed", confidence: <N/10>}.

1. If auto-invoked (from outclaw-plan): do ONE channel (the one

requested), silent, then return to the caller. NEVER ask the user for "preferred tone" or "desired format" — the whole point of this skill is that we infer style from samples. If a channel has no outbound samples at all (e.g. user just connected Discord today), log an observation: {type: observation, key: "style_nosamples_<channel>", insight: "no outbound samples on <channel>; using neutral template", source: "observed", confidence: 6} and emit a minimal neutral template at styles/<tenant>/<channel>_style.md with sample_count: 0, best_score: null, and a generic "direct, concise, warm-professional" prompt. Return to caller.

1. If user-invoked: show a compact report — which channels were

trained, best score each, sample count, confidence.

Filing rules (RESOLVER-compliant)

• Style prompt: styles/<tenant>/<channel>_style.md — NEVER

kb/styles/*, NEVER kb/me/styles/*.

• Raw outbound samples (the training data): kb/raw/style-<tenant>-<channel>-<ts>.jsonl.
• Summary for fast lookup: tenant memory, type=user,

key=style_trained_<channel>.

• Do NOT put learned styles in kb/me/self.md — voice description there

is for context, the style prompt lives separately.

Consent

Sample collection is a one-time opt-in, captured during outclaw-setup Step 2 as a memory preference entry. If no consent entry exists when this skill runs:

• Explicit invocation: ask once, record the decision.
• Auto-invocation from plan: proceed only if the tenant has an opt-in

style_consent entry. If missing, log an observation and fall back to a neutral template style. Never silently scrape without consent.

Output format (style prompt file)

---
tenant: outclaw
channel: gmail
trained_at: 2026-04-22T12:00:00Z
sample_count: 42
best_score: 82
best_iteration: 3
dimensions: [sentence_length, formality, personalization, cta_style, structure, tone, greeting_pattern, signoff_pattern]
conformity_log:
  - {iter: 1, score: 64}
  - {iter: 2, score: 71}
  - {iter: 3, score: 82}
  - {iter: 4, score: 80}
  - {iter: 5, score: 78}
---

# Gmail style — outclaw (tenant)

## Instructions
<the learned prompt — 500-2000 words, direct + actionable>

## Avoid
- <anti-pattern 1>
- <anti-pattern 2>

## Reference samples
- <path>/raw/style-outclaw-gmail-<ts>.jsonl (not included verbatim here;
  pointer only)

This format lets outclaw-plan parse the "Instructions" section directly into the draft-generation prompt.