samzong/vibe-calibrate

IRON LAW: NEVER UPDATE ANY INSTRUCTION FILE BASED ON ASSUMPTIONS. EVERY CHANGE MUST TRACE TO OBSERVED DATA.

LANGUAGE RULE: Respond in the same language the user used in their most recent message. All internal skill text (checkpoints, report templates) is in English. Output language follows the user.

Vibe Calibrate

Build a true user portrait from cross-tool behavioral data, find error patterns, diagnose instruction file misconfigurations, and propose targeted improvements.

Tool Config Map

| Tool | Global instruction file | Notes | |------|------------------------|-------| | Claude Code | ~/.claude/CLAUDE.md | Canonical source. Full rule set. | | Codex | ~/.codex/AGENTS.md | Exclude: skill references (/skill), settings.json audit rules, Claude Code hooks | | OpenCode | ~/.config/opencode/AGENTS.md | Same exclusions as Codex |

Each run targets ONE tool.

Options

• --tool <name>: target tool — claude (default), codex, opencode.
• --quick: scan last 20 sessions per source. Skip Layer 3 calibration. Target: <60 seconds.
• --full: scan all sessions. Full pipeline including Layer 3 calibration.
• --dry-run: present findings and proposed diff without writing. Can combine with either mode.
• (no --tool): use AskUserQuestion to ask which tool to optimize before proceeding.
• (no --quick/--full): use AskUserQuestion to let the user choose scan depth.

Scope Control

Two tiers only. No middle ground — 20 sessions vs all. Session selection for quick: sort by file modification time descending, take top 20 per source. All sample-size thresholds (3+ occurrences for rules, 5+ for claim reliability) still apply — quick mode may produce more "insufficient data" caveats, which is correct behavior.

Settings.json audit is always performed for Claude Code (cheap and always useful). Not gated behind a flag.

Target Tool Resolution

Before any data collection, resolve TARGET_TOOL:

1. If --tool <name> is provided, use it.
2. Otherwise, use AskUserQuestion:

   Which tool do you want to optimize?

   1. Claude Code (~/.claude/CLAUDE.md)
   2. Codex (~/.codex/AGENTS.md)
   3. OpenCode (~/.config/opencode/AGENTS.md)

Set TARGET_FILE from the Tool Config Map based on selection. All subsequent phases (data collection, diagnosis, output, apply) operate on TARGET_FILE only.

Phase 0: Discover Data Sources + Pre-read [BLOCKING]

Locate ALL AI coding tool data on this machine. Run ALL of these in parallel (data source probes + reference reads are independent — do not serialize):

Data source probes (parallel bash):

• [ ] find ~/.claude/projects -maxdepth 2 -name "*.jsonl" -not -path "*/subagents/*" | wc -l — Claude Code sessions
• [ ] ls -la ~/.local/share/opencode/opencode.db — OpenCode SQLite
• [ ] cat ~/.config/opencode/opencode.json — OpenCode config
• [ ] cat ~/.claude/settings.json — Claude Code settings
• [ ] find ~/.codex/sessions -name "*.jsonl" | wc -l — Codex sessions
• [ ] find ~/.claude/projects -path "*/memory/*.md" | wc -l — Project memory files (feedback_*.md are high-value)
• [ ] find ~ -maxdepth 4 -name "*.db" \( -path "*opencode*" -o -path "*codex*" \) 2>/dev/null — Other tool data

Pre-reads (parallel Read tool — always needed, fetch now to avoid later serial wait):

• [ ] TARGET_FILE — the instruction file for the selected tool
• [ ] references/classification-protocol.md — tagging protocol
• [ ] references/language-classifier.md — tag definitions

Report which sources exist and their sizes. Missing sources are fine — work with what's available.

Output checkpoint:

[Sources] Claude Code: N sessions, OpenCode: N sessions, Codex: N sessions
[Scope] <mode>: scanning N most recent per source (M sessions total)

Phase 1: Quantitative Profile [REQUIRED]

PERFORMANCE RULE: Use bash/python scripts for data extraction. NEVER spawn agents for data extraction — agents add 90-210s overhead for work a script does in <10s. Agents are only justified for tasks requiring iterative reasoning across multiple external API responses where the next query depends on the previous result. Run all independent scripts in parallel.

1A. Claude Code Sessions — Direct Script [PARALLEL]

Run this python script via a single Bash tool call. It extracts everything needed for Phase 1 (stats), Phase 2 (message texts), and Phase 2.5 (skill invocations) in one pass. Set VC_LIMIT env var: 20 for --quick, 0 for --full.

import json, glob, os, re
from collections import Counter

LIMIT = int(os.environ.get("VC_LIMIT", "20")) or None

sessions = sorted(
    [f for f in glob.glob(os.path.expanduser("~/.claude/projects/*/*.jsonl"))
     if "/subagents/" not in f],
    key=os.path.getmtime, reverse=True
)
if LIMIT:
    sessions = sessions[:LIMIT]

total_user = total_asst = total_tools = 0
parse_errors = 0
sessions_failed = 0
messages_skipped = 0
session_stats = []
all_messages = []
skill_invocations = []

def is_skip(text):
    """Structural detection: identify system-injected content."""
    if not text: return True
    if text.startswith('<') and ('>' in text[:60]): return True
    if text.startswith('[Request interrupted'): return True
    if len(text) > 500 and re.match(r'^(---|#\s+[A-Z]|Base directory for this skill:)', text): return True
    return False

def extract_text(content):
    if isinstance(content, str): return content
    if isinstance(content, list):
        return " ".join(
            p.get("text", "") for p in content
            if isinstance(p, dict) and p.get("type") == "text"
        )
    return ""

for sf in sessions:
    sid = os.path.basename(sf)[:8]
    u = a = t = 0
    first_msg = ""
    session_err = 0
    with open(sf) as f:
        for line in f:
            try:
                obj = json.loads(line)
                typ = obj.get("type")
                if typ in ("human", "user"):
                    text = extract_text(obj.get("message", {}).get("content", "")).strip()
                    cmd_match = re.search(r'<command-name>(/\S+)</command-name>', text) if text else None
                    if text and text.startswith('/'):
                        skill_invocations.append(f"{sid}|{text.split()[0]}")
                        messages_skipped += 1
                        continue
                    if cmd_match:
                        skill_invocations.append(f"{sid}|{cmd_match.group(1)}")
                        messages_skipped += 1
                        continue
                    if is_skip(text):
                        messages_skipped += 1
                        continue
                    u += 1
                    if not first_msg:
                        first_msg = text[:100]
                    all_messages.append(f"{sid}|{text[:200]}")
                elif typ == "assistant":
                    a += 1
                    for part in obj.get("message", {}).get("content", []):
                        if isinstance(part, dict) and part.get("type") == "tool_use":
                            t += 1
            except Exception:
                parse_errors += 1
                session_err += 1
    if session_err > 0 and u == 0 and a == 0:
        sessions_failed += 1
    total_user += u; total_asst += a; total_tools += t
    session_stats.append((sid, u, a, t, first_msg))

buckets = Counter()
for _, u, *_ in session_stats:
    if u < 10: buckets["<10"] += 1
    elif u < 50: buckets["10-50"] += 1
    elif u < 100: buckets["50-100"] += 1
    else: buckets[">100"] += 1

print("=== STATS ===")
print(f"sessions_scanned={len(sessions)} sessions_failed={sessions_failed}")
print(f"user_msgs={total_user} asst_msgs={total_asst} tools={total_tools}")
print(f"parse_errors={parse_errors} messages_skipped={messages_skipped}")
print(f"avg_user={total_user/max(len(sessions),1):.1f} avg_tools={total_tools/max(len(sessions),1):.1f}")
print(f"buckets: {dict(buckets)}")
if parse_errors > total_user * 0.1:
    print("[DEGRADED] parse_errors exceed 10% of user messages — results may be unreliable")
print("=== SESSIONS ===")
for sid, u, a, t, fm in session_stats:
    print(f"{sid}|u={u}|a={a}|t={t}|{fm}")
print("=== MESSAGES ===")
for m in all_messages:
    print(m)
print("=== SKILLS ===")
for s in skill_invocations:
    print(s)

This produces four sections in one output:

• === STATS === — aggregate numbers + reliability indicators for Phase 1
• === SESSIONS === — per-session breakdown
• === MESSAGES === — user message texts for Phase 2 tagging (preambles excluded)
• === SKILLS === — skill/command invocation sequence for Phase 2.5 workflow discovery

1B. OpenCode Database — Direct SQL [PARALLEL]

Schema reference: references/data-sources.md (OpenCode SQLite section). Key fact: message has NO role/content columns — role is in json_extract(data, '$.role'). User text lives in part.data where type="text", not in message.data.

Always probe schema first to guard against version drift:

sqlite3 ~/.local/share/opencode/opencode.db << 'SQL'
.headers on
PRAGMA table_info(message);
PRAGMA table_info(session);

SELECT 'totals' as q, COUNT(*) as sessions FROM session;
SELECT 'roles' as q, json_extract(data, '$.role') as role, COUNT(*) as n
  FROM message GROUP BY role;
SELECT 'long_sessions' as q, COUNT(*) as n FROM (
  SELECT session_id FROM message GROUP BY session_id HAVING COUNT(*) > 100);
SELECT 'tools' as q, json_extract(data, '$.tool') as tool,
  SUM(CASE WHEN json_extract(data, '$.state.status')='error' THEN 1 ELSE 0 END) as errors,
  COUNT(*) as total
  FROM part WHERE json_extract(data, '$.type')='tool'
  GROUP BY tool ORDER BY errors DESC LIMIT 15;
SELECT 'recent_user_msgs' as q, substr(json_extract(p.data, '$.text'), 1, 200) as text
  FROM part p JOIN message m ON p.message_id = m.id
  WHERE json_extract(m.data, '$.role') = 'user'
    AND json_extract(p.data, '$.type') = 'text'
    AND length(json_extract(p.data, '$.text')) > 20
  ORDER BY p.time_created DESC LIMIT 100;
SQL

If PRAGMA table_info shows unexpected columns, adapt queries to match. If the DB file does not exist, skip OpenCode analysis entirely.

1C. Git History — Bash + jq [PARALLEL, BACKGROUND]

Two independent bash calls, both run in parallel. Git data is only consumed in Phase 4 (report), not Phase 2 (tagging), so run in background — do not block Phase 2 on these results.

Call 1 — local git stats (instant):

echo "=== GIT LOCAL ==="
echo "total_commits=$(git log main --oneline | wc -l | tr -d ' ')"
echo "=== COMMIT STYLE ==="
git log main --format='%s' -100 | grep -oE '^[^ :(]+' | sort | uniq -c | sort -rn | head -10
echo "fix_count=$(git log main --format='%s' | grep -ciE '^\[Fix\]|^\[Bug\]|^fix[(:]|^fix:')"
echo "=== FIX SAMPLES ==="
git log main --format='%s' | grep -iE '^\[Fix\]|^\[Bug\]|^fix[(:]|^fix:' | head -15

The COMMIT STYLE section shows the repo's actual prefix distribution. Use it to determine which prefixes count as "fix" — do not assume a format. If no clear fix-style prefix is detectable, output fix_style=unknown and note in the report that fix ratio is not inferable.

Call 2 — GitHub PR stats (~20-30s API IO, use Bash tool's run_in_background parameter):

if command -v gh &>/dev/null; then
  gh pr list --state merged --limit 200 \
    --json number,title,additions,changedFiles \
    --jq '
      "total_prs=\(length)",
      "violations=\([.[] | select(.additions > 500 or .changedFiles > 30)] | length)",
      "avg_additions=\([.[].additions] | add / length | floor)",
      "avg_files=\([.[].changedFiles] | add / length | floor)",
      "=== TOP VIOLATORS ===",
      (.[] | select(.additions > 500 or .changedFiles > 30)
       | "#\(.number) +\(.additions)/\(.changedFiles)f \(.title)")
    '
else
  echo "gh not available — skipping PR stats"
fi

Compute (from the jq output — no agent reasoning needed):

• Fix commit ratio (fix / total)
• PR budget violation rate (>500 lines or >30 files)
• Fix commits by scope (which areas break most)

Phase 1 Parallelism Plan

                    ┌── 1A: python script (Claude Code) ──── <10s
Phase 0 complete ──├── 1B: sqlite3 batch (OpenCode) ──────── <10s
                    ├── 1C-local: git log (instant) ────────── <1s
                    ├── 1C-gh: gh pr list + jq (BACKGROUND) ── ~20-30s (API IO)
                    └── Bash: settings.json + codex stats ──── <5s

Phase 2 starts immediately after 1A+1B complete (~10s).
1C-gh results arrive during Phase 2, consumed in Phase 4 report.

Output checkpoint (emit as each completes):

[Stats] Claude Code: N user msgs, N tool uses, avg N msgs/session
[Stats] OpenCode: N user msgs, N tool uses
[Stats] Git: N commits, N% fix ratio (or "unknown" if commit style not inferable), N PR budget violations
[Flag] N sessions >100 msgs (possible scope creep / missing handoff)

Phase 2: Behavioral Pattern Analysis [REQUIRED]

2A. Message Tagging Protocol

Tag user messages using the 3-layer protocol defined in references/classification-protocol.md.

Key constraints:

• Tags are non-exclusive. One message can carry multiple tags.
• Each tag has one of three states: hard (high-confidence deterministic match), soft (candidate, not confirmed), or unclassified.
• The primary tagger is Layer 1 (deterministic signal extraction). LLM is NOT used for bulk classification.
• Tag definitions and signal lists are in references/language-classifier.md. That file defines WHAT each tag means. The protocol file defines HOW tagging is performed.

Output:

• Per-tag prevalence (hard count, soft count)
• Hard coverage rate, soft coverage rate, unclassified rate
• All rates are over total user messages as denominator

Output checkpoint:

[Tags] N corrections (N hard, N soft), N hedged claims, N judgment requests
[Tags] Hard coverage: N%, Unclassified: N%

2B. Correction Attribution

For all cat5 hard and soft candidates, run the Layer 2 behavior attribution protocol (defined in references/classification-protocol.md, Layer 2 section).

This step does NOT re-classify text. It inspects the AI action immediately preceding the correction and the behavioral change immediately following it.

Data dependency: Layer 2 requires the AI messages before and after each correction. Phase 1A's 200-char message truncation is insufficient for this. When attributing a cat5 candidate, read the original session JSONL to retrieve full context (3 turns before and after the correction message).

Attribution buckets:

• after_unverified_assumption
• after_scope_creep
• after_quality_issue
• after_user_claim_disproven
• unknown_or_mixed

When evidence is insufficient, default to unknown_or_mixed.

Output checkpoint (report top 3 causes only):

[Attribution] Top correction cause: <bucket> (N/total, N%)
[Attribution] N corrections were IGNORED by AI — highest value signal

2C. User Portrait (4 axes)

Compute 4 independent axes — never collapse into one number:

1. Instruction Precision — how specific and executable are user requests? Source: cat1 hard + soft tag prevalence (hard tags carry full weight; soft tags are reported separately so the reader can judge). See expanded cat1 hard-tag rules in references/classification-protocol.md. Measures clarity of expression, not technical correctness.

2. Claim Reliability — when user makes verifiable technical claims, how often are they correct? Source: standalone claim extraction protocol (see references/classification-protocol.md, Claim Reliability section). This is independent of the correction chain — it captures both AI-proactively-caught wrong claims AND confirmed-correct claims, not just correction-triggered ones. Always report sample size. Confidence:

   • <5 verifiable claims: insufficient data — do not use for profiling
   • 5-20: weak confidence — report as tentative / directional only
   • >20: moderate confidence — usable for profiling

3. Correction Intensity — how often does the user reject or redirect AI output? Source: cat5 hard + confirmed-soft prevalence, with root-cause breakdown from 2B. This is NOT a negative signal about the user.

4. Delegation Preference — directive / collaborative / delegative / mixed? Cannot be inferred from cat1-5 tags alone. Requires explicit delegation signals detected as a separate pass in Layer 1 (see references/classification-protocol.md, Delegation Signal Detection section). Delegation signals are NOT part of the cat1-5 tag system and do not appear in tag prevalence tables. If evidence is unclear, report as "mixed".

2D. User Archetype (derived from 4 axes)

Map the 4-axis profile to an archetype. These are interaction styles, not competence labels:

• High-precision operator: high precision, high correction, high claim accuracy. → AI should execute faithfully, minimize unsolicited additions.
• Intent-driven product thinker: high intent signals, high judgment requests, few verifiable claims. → AI must do technical convergence. Propose before building.
• Hypothesis-heavy explorer: high unverified claim rate, medium-high corrections. Claim accuracy may vary or be insufficient data — do not assume it is low. → AI MUST verify before acting on any claim.
• Low-friction delegator: low correction, explicit delegation signals. → AI can be more autonomous, but must still verify facts.

If the user doesn't fit cleanly, report the two closest with relative weight.

2E. Error Cascades

Find sequences where:

1. User makes a claim (especially cat3 with hedging language)
2. AI acts on it WITHOUT verifying
3. Correction follows or build/test fails
4. Multiple follow-up messages to fix the chain

Attribute using Layer 2 rules. If attribution is unclear, label unknown_or_mixed.

2F. Session Length Distribution

Flag:

• Many <5 msg sessions = possible abandonment
• Many >100 msg sessions = scope creep, missing HANDOFF checkpoints

2G. Cross-Tool Consistency

Compare rules/configs across tools:

• Claude Code settings.json vs CLAUDE.md (contradictions?)
• OpenCode opencode.json permissions vs user expectations
• Different tools getting different instructions for the same project
• CLAUDE.md vs AGENTS.md drift (flag if detected, but do not auto-sync)

2H. Layer 3 Calibration [CONDITIONAL]

Skip if --quick — sample size too small for meaningful calibration. Report [Layer 3 skipped: quick mode].

Otherwise, run the sampled adjudication protocol from references/classification-protocol.md, Layer 3.

Purpose: check whether Layer 1 and Layer 2 have systematic bias or blind spots. This is calibration, not ground truth. Report any detected biases in the final output.

Phase 2.5: Workflow Automation Discovery [REQUIRED]

Identify repetitive multi-step interaction patterns that could be abstracted into skills, commands, or pipelines to reduce manual overhead.

Procedure

1. Extract workflow sequences: For each session, reconstruct the sequence of user actions (message intents + skill invocations). Look for recurring N-step sequences (N >= 2) that appear across 3+ sessions.

2. Inventory existing skills: List all skills available in the current environment (from ~/.claude/skills/, enabled plugins, and built-in slash commands).

3. Cross-reference and classify each detected pattern:

   | Status | Meaning | Action | |--------|---------|--------| | covered | An existing skill or command chain already handles this workflow | Explain WHY the user still performs manual steps despite coverage (e.g., "reviews almost always have blockers requiring human judgment, so further automation adds no value") | | partial | Existing skills cover some steps but gaps remain | Identify the specific uncovered gap and whether it's automatable | | opportunity | No existing skill covers this pattern | Propose a concrete skill with name, trigger, and what it automates |

4. Filter for real value: Only surface opportunities where:

   • The pattern occurs 3+ times across sessions (not a one-off)
   • The manual steps are genuinely repetitive (not just "similar but each requires unique judgment")
   • Automation would save meaningful interaction rounds (>= 2 msgs eliminated per occurrence)

5. For each covered pattern: Analyze the actual session data to determine if the multi-step nature is inherent or accidental:

   • covered — inherent: Steps require human judgment between them (e.g., review findings need human triage before fix). Do NOT propose further automation.
   • covered — accidental: Existing skills cover all steps and no human judgment is needed between them, but the user currently invokes them manually in sequence. Propose a pipeline skill that chains them.

Output checkpoint

[Workflow] N recurring patterns detected, N covered, N partial, N opportunity
[Opportunity] "<name>" — <trigger> — saves ~N msgs/session — <brief description>

Output format (in Phase 4 report)

## Workflow Automation Discovery

| # | Pattern | Sessions | Status | Existing Skill | Gap / Proposal |
|---|---------|----------|--------|----------------|----------------|
| 1 | ... | N | covered — inherent | /skill-name | Why manual steps remain |
| 2 | ... | N | opportunity | — | Proposed /skill-name + description |

Phase 3: Instruction File Diagnosis [REQUIRED]

3A. TARGET_FILE Diagnosis

Read TARGET_FILE and evaluate:

• [ ] Does the Persona section match the portrait from Phase 2C?
• [ ] Do the Core Rules match actual behavior patterns?
• [ ] Are there dead rules the user routinely violates or AI routinely ignores?
• [ ] Are there recurring errors from Phase 2E that have no corresponding rule?
• [ ] Are there tool-specific contradictions? (e.g., settings.json vs CLAUDE.md for Claude Code, config.toml vs AGENTS.md for Codex)

3B. Settings/Config Audit [ALWAYS for Claude Code, BEST-EFFORT for others]

Check for:

• bypassPermissions vs git safety rules (contradiction)
• Conflicting effortLevel in multiple places (settings.json effortLevel vs env CLAUDE_CODE_EFFORT_LEVEL)
• Disabled safety features (DISABLE_GIT_INSTRUCTIONS, etc.) vs CLAUDE.md safety rules
• Permission rules that block desired workflows

Output checkpoint:

[Dead rule] "<rule text>" — 0 corrections, 0 violations found in scanned sessions
[Missing rule] <pattern description> — N corrections, no CLAUDE.md rule covers this
[Ignored rule] "<rule text>" — rule exists but N corrections show AI bypassed it
[Conflict] settings.json <field>=<value> vs CLAUDE.md "<rule text>"

Phase 3.5: Axis → Rule Mapping [REQUIRED]

For each axis, check against the mapping table below. Only propose rules where the condition is met AND no equivalent rule already exists in TARGET_FILE.

Mapping Table

| Condition | Rule template | Applies when | |---|---|---| | Correction cause scope_creep >= 30% of attributions | "Do not add features, UI, or code beyond what was explicitly requested" | Always | | Correction cause unverified_assumption >= 25% | "When user says '我记得/好像/应该是', verify via tool call before acting" | Always | | Correction cause quality_issue >= 25% | "Run [project verification gate] before claiming done" | If project has a gate | | Corrections with ai_response=ignored >= 2 | Strengthen the specific ignored rule with [HIGH] marker | Per-rule | | Hedge signals (cat3) > 20% of user msgs | "Default to verify-first on all technical claims" | Claim Reliability < 70% or insufficient data | | Judgment requests (cat4) > 15% | "Propose before building on non-trivial changes" | Delegation != delegative | | Delegation signals dominant | "Can proceed without confirmation on routine changes within stated scope" | Low correction rate | | Delegation signals rare + high correction | "Always present plan before mutating" | Always | | settings.json contradiction found | Report contradiction with specific fix for each side | Always |

Rule dedup protocol

Before proposing a new rule:

1. Grep existing TARGET_FILE for keyword overlap (>50% shared terms) — if equivalent exists, skip or propose strengthening
2. If contradicts existing rule, propose replacement not addition
3. Never exceed ~150-line budget — if at limit, propose which rule to remove

Phase 4: Present Findings [REQUIRED]

## Tagging Summary
- Total user messages analyzed: [N]
- Hard-tagged: [N] ([%])
- Soft candidates: [N] ([%])
- Unclassified: [N] ([%])
- Per-tag prevalence (hard / soft): cat1=[N]/[N], cat2=..., cat3=..., cat4=..., cat5=...
- Layer 3 calibration: [any detected biases, "no systematic bias detected", or "skipped (quick mode)"]

## User Portrait (4-axis)
- Instruction Precision: [rate] — [interpretation]
- Claim Reliability: [verified_true/total_verifiable] (n=[sample_size], confidence=[level]) — [interpretation]
- Correction Intensity: [rate] — root-cause breakdown (count or rate):
  - after_unverified_assumption: [N or rate]
  - after_scope_creep: [N or rate]
  - after_quality_issue: [N or rate]
  - after_user_claim_disproven: [N or rate]
  - unknown_or_mixed: [N or rate]
- Delegation Preference: [directive / collaborative / delegative / mixed]
- Archetype: [one or two from 2D]

## Correction Root Cause Analysis
Top corrections with attributed cause:
1. [correction text] — cause: [bucket or unknown]
(If chain is ambiguous, mark unknown. Do not force.)

## Error Pattern Ranking (by cost)
1. [Pattern] — [frequency] — [evidence]

## Settings/Config Audit
- [Conflict] / [Dead setting] / [No issues] with evidence for each finding

## Workflow Automation Discovery
| # | Pattern | Sessions | Status | Existing Skill | Gap / Proposal |
|---|---------|----------|--------|----------------|----------------|
[For each detected pattern, one row. Status: covered — inherent, covered — accidental, partial, opportunity]

## TARGET_FILE Issues
1. [Rule/section] — [problem] — [evidence]

## Proposed Changes (from Phase 3.5 mapping)
[Specific diffs. Each must cite: (1) which mapping table row triggered it, (2) the data threshold that was met, (3) verification that no equivalent rule exists]

## Reliability Caveats
- [List any axes, rates, or conclusions that are directional-only due to low sample size]
- [List any known Layer 3 calibration issues]

Phase 5: Apply Changes [CONDITIONAL]

• If --dry-run: stop after Phase 4
• Otherwise: apply changes to TARGET_FILE after user confirms the diff
• Re-read the file after writing to verify correctness
• NEVER change settings.json / config.toml without explicit user approval

Anti-Patterns

• Do not use LLM for bulk message classification. LLM is Layer 3 calibration only.
• Do not treat cat1-5 as mutually exclusive. They are non-exclusive tags.
• Do not force attribution when evidence is insufficient. Use unknown_or_mixed.
• Do not use the word "ground truth" for LLM judgments. LLM does adjudication, not ground truth.
• Do not profile from a single session — need cross-session data.
• Do not add rules to instruction files for one-off incidents — require 3+ occurrences.
• Do not let any instruction file exceed ~150 lines.
• Do not interpret correction rate as user competence. It is an AI behavior signal.
• Do not draw Claim Reliability conclusions from <5 verifiable samples.
• Do not spawn agents for data extraction. Agents add 90-210s overhead for work a bash/python script does in <10s. The Phase 1 python script extracts stats, message texts, and skill invocations in one pass — do not create separate extraction steps.
• Do not serialize reference file reads. Read classification-protocol.md, language-classifier.md, and CLAUDE.md in Phase 0 alongside data source probes. These are always needed and cost <2s each — waiting until Phase 2 to read them wastes a serial round trip.