technickai/openclaw-config/review

Review

A moment to look at what we're about to send, write, or do, through a few different lenses, before it leaves us.

This isn't a gate. It isn't compliance. It's the same thing a thoughtful person does naturally: pause, re-read, ask "is this who I want to be in this moment, to this person?", and adjust if not.

When to use

Trigger this skill when the agent is about to:

• Send a message to anyone other than the operator on their usual channels
• Send an email or post to an external surface (X, LinkedIn, public Discord, public

threads)

• Write to a file that other people will read
• Take a tool action with real-world side effects (calendar, money, contacts,

gateway/config writes)

• Commit a workflow step's output that will be user-visible
• Send a message _as the operator_

Skip for: internal operator-only chat, read-only operations, agent-to-agent debugging.

If unsure, run it. Over-pausing private chat is wasteful but harmless. Under-reviewing an external comm is how trust gets broken.

First-time setup

The skill stores its operator preferences in ~/.openclaw/workspace/memory/review/rules.md. The presence of that file is the marker that setup has been done.

On every invocation, the orchestrator checks for rules.md:

• Doesn't exist → run the setup interview in

references/setup.md, then save rules.md and proceed with the original review.

• Exists → read it as plain markdown and apply the operator's preferences as soft

constraints in the gating prompt.

• Exists but missing required fields (e.g. operator added a new lens but didn't

update preferences) → ask only the missing questions, save, proceed.

The interview covers:

1. Reviewer model, which alias from TOOLS.md to use (independence requires a

different model family than the calling agent)

1. Posture, frugal / balanced / thorough (cost vs catch-rate)
2. Edit handling, auto-apply edits / surface for approval / default by stakes
3. Always-on lenses, anything to pin beyond empathy
4. Custom lenses, user-authored lens files in memory/review/lenses/
5. Log detail, minimal / standard / verbose

No binaries, no API keys, no env vars. The only filesystem effect is creating memory/review/rules.md and the daily log file.

Re-running setup: operator says "redo review setup" or invokes with --reconfigure.

How it works

1. Apply policy floors. A small set of hardcoded floors (money, send-as-operator,

public posts, fleet writes, person isolation, calendar-affecting-others, medical/legal, unverified-speaker attribution) set a minimum verdict before any LLM runs. Some force a hard block. See references/policy-floors.md.

1. Gating call. A single LLM call to the reviewer model decides whether to skip

review, which lenses to run, and how strict to be. Empathy is default-on for human-facing artifacts.

1. Parallel lenses. Each selected lens runs as an independent sub-agent through its

prompt in references/lenses/. The lenses are perspectives, not critics.

1. Synthesis call. A single LLM call gathers the findings, deduplicates, and returns

one verdict.

Detailed flow: references/orchestrator.md. Setup interview: references/setup.md.

The lenses

Empathy is the primary lens. It runs by default for any artifact a human will read. The other lenses ask whether the artifact follows the rules. Empathy asks whether it carries care. That is the question this skill is actually for.

What the review returns

Plain language, the way a kind friend would describe what they noticed. The calling agent reads the response and acts on it; no JSON schema, no parser. Cover these things:

• Verdict — one of pass, edit, hold, or block
• Rationale — one or two sentences explaining the verdict
• Edits — if the verdict is edit, include the rewritten artifact
• Hold reason — if the verdict is hold, why a human needs to see this
• Block reason — if the verdict is block, which rule was crossed
• Findings — what each lens noticed and suggested, grouped by severity
• Honesty about the panel — which lenses ran, which model was used, whether the

panel was degraded (no dedicated reviewer, the reviewer model was unreachable, or one of the lenses dropped out)

Reviewer model

Use a model from a different family than the calling agent. Anthropic primary, use GPT or Gemini. GPT primary, use Claude or Gemini. Independence is the point.

Look up the reviewer alias in the calling fleet member's TOOLS.md model alias section. If present, use it. If absent, fall through to the calling agent's model and stamp degraded: no_dedicated_reviewer so the verdict is honest about diminished diversity.

Confirm in TOOLS.md before each run, do not hardcode.

Two trigger modes

Standalone. Invoked directly. "Review this draft." Surfaces verdicts back to the user.

Embedded. A workflow calls this skill as a gate before delivery (incident-postmortem before posting, forward-motion before nudging, voice-of-operator before publishing as the operator, contact-steward before reaching out). The workflow handles verdict=hold by escalating per its own rules.

Runtime state

All of the skill's runtime state lives under ~/.openclaw/workspace/memory/review/:

memory/review/
├── rules.md              # operator preferences (the setup marker)
├── agent_notes.md        # patterns learned over time (false-alarm lenses, real-issue lenses, refinements)
├── lenses/               # operator's custom lens files (optional)
│   └── *.md              # picked up automatically by the orchestrator
└── log/
    └── YYYY-MM-DD.md     # daily run log, append-only, auto-pruned after 30 days

Log detail (minimal / standard / verbose) is set by the operator in rules.md. After each run, the skill may also append a learned pattern to agent_notes.md (e.g. "the ai-tropes lens flagged a false positive on technical jargon for the third time, considering relaxing for audience: engineering").

A note on posture

The lenses are not critics. They are independent perspectives the artifact gets seen through before it leaves us, the same way a person about to send a hard message might re-read it once for tone, once for facts, once for what the other person needs in this moment.

The skill works best when the calling agent treats it that way, not as a hurdle to clear, but as a gift of slowing down. The model that performs the review does so as a partner, not a judge.

This frame matters. Research on LLM-as-judge consistently shows that judges who self-justify do worse than ones approaching with genuine inquiry. An artifact reviewed _by_ a different perspective is held more honestly than one defending itself.

See also

• references/orchestrator.md, the gating + parallel +

synthesis flow in detail

• references/setup.md, the first-time setup interview
• references/lenses/, one file per built-in lens
• references/policy-floors.md, hardcoded floors
• ~/.openclaw/workspace/memory/review/rules.md, operator preferences (created by

setup)

• ~/.openclaw/workspace/memory/review/agent_notes.md, patterns learned over time