run_meta_critic

run_meta_critic

This skill is a thin wrapper that dispatches the meta-critic Claude Code subagent (defined in ${CLAUDE_PLUGIN_ROOT}/agents/meta-critic.md) against a specified analysis directory. The agent does the work; this skill just resolves inputs and surfaces the verdict.

When to use this vs. the automatic invocation

Situation	Use
Running the full pipeline	The orchestrator handles meta-critic in its loop — no need to call this skill
Auditing analyses that were produced outside the orchestrator	This skill
Re-auditing after manually editing an analyst report	This skill
Cross-checking historical analyses against an updated meta-critic.md prompt	This skill

Inputs

When invoked, ask the user (or accept arguments) for:

Input	Description	Default
analyses_dir	Directory containing the analyst reports to audit	the most recent session under $HOME/.claude/agent-twin/personalized/saves/session/
session_id	The session ID for traceability	inferred from analyses_dir (parent directory name)
iteration	Iteration number — meta-critic enforces escalation at iteration 3	1 if running standalone

The skill auto-resolves:

• agent_prompts_dir → ${CLAUDE_PLUGIN_ROOT}/agents/
• source_data_path → <session-dir>/annotated.txt
• synthesis_path → <analyses-dir>/synthesis.md if it exists, else empty

Step 1 — Resolve inputs

If the user did not specify analyses_dir, list $HOME/.claude/agent-twin/personalized/saves/session/ and pick the most recently modified subdirectory's analyses/. Confirm the choice with the user before proceeding.

Verify required files exist:

• At least the four analyst reports (affect.md, social.md, values.md, narrative.md)
• The source data (annotated.txt)

If files are missing, list what is missing and stop.

Step 2 — (Optional) review the meta-critic prompt

The agent file ${CLAUDE_PLUGIN_ROOT}/agents/meta-critic.md is loaded automatically by Claude Code when you dispatch the meta-critic subagent — you do not need to substitute variables into the prompt body yourself. Read it only if you want to confirm the audit dimensions or its output-contract block before dispatching.

Step 3 — Dispatch the agent

Issue a Task call with subagent_type: "meta-critic" and a task prompt that lists these input variables (Claude Code loads the system prompt from the agent file automatically — do not re-include it):

Variable	Value
ANALYSES_DIR	resolved above
AGENT_PROMPTS_DIR	${CLAUDE_PLUGIN_ROOT}/agents/
SOURCE_DATA_PATH	<session-dir>/annotated.txt
SYNTHESIS_PATH	<analyses-dir>/synthesis.md if exists, else empty
SESSION_ID	resolved above
ITERATION	as specified, default 1

Wait for completion.

Step 4 — Save the audit report

Save the agent's output to <analyses-dir>/meta_critic.md. If a previous meta_critic.md exists, rename the prior one to meta_critic.<timestamp>.md before overwriting (manual audit reports are kept as history; only the orchestrator-driven loop overwrites freely).

Step 5 — Surface the verdict to the user

Read the saved report and surface the key findings:

1. Loop decision — accept / iterate / escalate
2. Per-analyst verdicts — pass / pass_with_warnings / needs_revision / unrecoverable
3. Top concerns — anchoring residual rate, blind spots, confidence-calibration flags
4. Path to the full audit report

Step 6 — Recommend next action

Based on the verdict:

Verdict	Recommendation
accept	"Phase 1 output is sound. The downstream phases can use this synthesis."
iterate (any analyst needs revision)	"These analysts need revision — re-run the full pipeline via /run_pipeline (the orchestrator handles the iteration loop) or hand-edit the analyst report and re-invoke this skill to re-audit."
escalate	"Pipeline cannot self-correct. Review the audit report's escalation notes — human judgment required."

Do not automatically re-dispatch analysts from this skill — that is the orchestrator's job. This skill is read-only with respect to analyst outputs.

Completion checklist

• Inputs were resolved (or asked of the user) before dispatching
• All required source files exist
• Agent dispatch used the correct variable substitution
• Prior meta_critic.md (if any) was preserved as meta_critic.<timestamp>.md
• The verdict was surfaced clearly to the user
• No analyst report was modified
• No re-dispatch of analysts (that is orchestrator's role)

Why this exists

The orchestrator is the right runner for the full pipeline; it handles iteration and synthesis automatically. But after a pipeline has finished, users sometimes want to re-audit — perhaps because the meta-critic prompt was updated, or because they hand-edited a report to fix something. Running the orchestrator from scratch in those cases is wasteful. This skill is the lightweight alternative.