CLAUDE-BRAIN · OPS LOG · VOL 01
2026.06.02 · CHARLOTTE NC
A Field Report
By Mahmoud "Michael" Kurdi
For the next instance of the agent
Issue Nº 001
Filed 02 · 06 · 2026
Commit · b68fa99
The Overhaul

The quiet war on
token bloat & silent drift.

A configuration archaeology — diagnosing the ghosts in ~/.claude/, exorcising what was loaded but never used, and rewiring the system so it observes itself, prunes itself, and gets sharper every Friday at 3pm.

System State · OPTIMISED
— 40%
Cold-Start Reduction
001
Chapter One · The Diagnosis

Nine ghosts
in the machine.

The visible symptom was "too many skills." The real disease was that almost every loaded surface was never invoked — only 4 of ~300 had fired across 150+ sessions. We named the ghosts before exorcising them.

i

Phantom plugins

31 enabled across 4 marketplaces. Each contributes skill metadata to the cold-start dump. Most never fired.

— 17 disabled
ii

Inline doppelgängers

Six plugin install instances were inline+marketplace duplicates of the same code, all empty stubs.

— 8 removed
iii

The orphan

plugin_011v5h6QUzBZvas64y44XLhy — no source, no manifest, no purpose. A leftover ID.

— deleted
iv

Skill cemetery

17 local skills authored over months. Telemetry says 4 ever fired. Thirteen were dead weight your context still paid for.

— 11 archived
v

Rule auto-load

Fifteen .md files in rules/ auto-loaded on every cold start — 34 KB before the user typed a single character.

— 100% lifted
vi

The dormant memory

MEMORY.md was three lines. No user/feedback/project/reference buckets. The protocol existed only on paper.

— wired live
vii

Twin hubs

Two CLAUDE.md files competed. The global one ordered the agent to immediately fetch a 13 KB second one. Doubled cost.

— unified
viii

Silent router

SKILL-ROUTER.md existed but was never loaded. Beautifully written documentation of a feature that didn't exist.

— activated
ix

No telemetry

Zero data on what was actually used. Every pruning decision was a guess. We were optimising a black box.

— observed
002
Chapter Two · The Intervention

Eleven pillars,
reversible-first.

Every destructive step was gated behind a backup. ~/.claude/ became a git repo before the first change — so rollback was always one command away.

The sequencing matters as much as the actions. Memory and router went in first because they're additive. Plugin flips came mid-sequence because they're reversible JSON. Disk deletes came last, gated by a Scout inspection that revealed all 8 orphan dirs were empty stubs — zero data loss, no real gate needed.

  1. i

    Backup foundation

    git init · .gitignore · baseline commit · settings backup
    done
  2. ii

    Rule mechanism scout

    Confirmed Claude Code auto-loads rules/*.md by convention
    done
  3. iii

    Rules-as-skill lift

    15 files moved to skills/rules/refs/ + dispatcher SKILL.md
    done
  4. iv

    Plugin prune

    17 of 33 enabled plugins flipped to false
    done
  5. v

    Disk cleanup

    8 empty stub plugin directories removed
    done
  6. vi

    Skill archive

    11 unused local skills moved to _archived/
    done
  7. vii

    Hub unification

    Global CLAUDE.md slimmed 121 → 73 lines; Brain-hub demoted
    done
  8. viii

    Router wiring

    Pattern→skill table now inline in CLAUDE.md
    done
  9. ix

    Memory protocol

    4-bucket dir + MEMORY.md index live
    done
  10. x

    Telemetry hooks

    Skill, prompt, session-end logging → telemetry/
    done
  11. xi

    Auto-evolution

    Weekly autonomous prune + auto router-sync on save
    done
003
Chapter Three · The Numbers

Measured,
not guessed.

Baseline captured before any change. Re-measured after the dust settled. Every line of strikethrough is a real ghost laid to rest.

Enabled plugins
33
16
−52%−17 surfaces
Auto-loaded rule bytes
34,022
0
−100%off cold start
Active local skills
17
7
−59%11 archived
CLAUDE.md size
7,468 B
4,457 B
−40%121 → 73 lines
Plugin install dirs
12
4
−67%stubs purged
Memory subdirs
0
4
+ protocol live
Telemetry hooks
0
4
+ self-observing
Estimated impact

≈ 10,000 tokens off every
cold start. Forever.

004
Chapter Four · The Machine

What's now
under the hood.

01 · IDENTITY

CLAUDE.md

Slim hub. Identity, stack defaults, execution rules, inline router. 73 lines, every one earning its place.

always loaded
02 · INDEX

MEMORY.md

Four-section index pointing to user, feedback, project, reference memory buckets.

always loaded
03 · DISPATCH

Skill router

Pattern→skill table inline in CLAUDE.md. Deterministic. Local skills self-register via auto-sync.

deterministic
04 · ON-DEMAND

rules dispatcher

The 15 conventions live here. Loads only the relevant file when keyword fires. Zero cold-start cost.

trigger-loaded
05 · MEMORY

memory/user/

Preferences, profile facts, corrections you want me to remember across sessions.

write on signal
06 · MEMORY

memory/feedback/

Session logs, prune reports, router suggestions, validated approaches. Self-improvement substrate.

auto-appended
07 · MEMORY

memory/project/

Per-project decisions, schemas, gotchas. Alexandria lives here.

manual write
08 · MEMORY

memory/reference/

External pointers: dashboards, repos, tools, useful patterns discovered in the wild.

manual write
09 · OBSERVE

telemetry/

Skill invocations, prompts, session ends — all logged. Substrate for the weekly auto-prune to reason against.

hooks · jsonl
10 · EVOLVE

auto-prune-weekly.sh

Fridays 3pm. Spawns Claude CLI. Archives 60+ day dead skills. Writes router suggestions.

windows task
11 · EVOLVE

sync-local-skill-router.js

New SKILL.md written → router table auto-updates between markers. Manual edits never overwritten.

PostToolUse + SessionStart
12 · SAFETY

git history

Three commits. Rollback = one command. Plugin re-enable = one JSON flip. Every step is reversible.

e0b4501 → b68fa99
005
Chapter Five · The Weekly Loop

A system
that prunes itself.

Every Friday at 3pm Eastern, Windows Task Scheduler wakes a one-shot Claude CLI agent. Bounded prompt, narrow surface, hard 15-minute cap. It reads its own telemetry and acts.

The agent will never delete a skill — only archive. It will never edit CLAUDE.md directly — only write suggestions to memory/feedback/router-suggestions-YYYY-MM-DD.md for your weekend review. It will never touch plugins, settings.json, or any project repo.

All changes are committed in ~/.claude/ with a conventional message. Rollback any prune run with git reset --hard HEAD~1.

Friday
15:00
Generate dead-skills report
Spawn bounded Claude CLI
Archive 60+ day dead skills
Propose router additions
Commit changes to git
Sleep until next Friday
ClaudeBrain-AutoPruneWeekly · 15min cap
006
Chapter Six · The Router

Speak the keyword,
load the skill.

Stop thinking about which skill to load. Phrase your request with the trigger keyword. The router does the rest. These are the most useful matches.

"threat-model the auth flow"
vulnerability · OWASP · RLS · SOC · NIST		security-scanning
"diagnose this race condition"
broken · throwing · failing · regression		diagnose · superpowers:systematic-debugging
"code review on this PR"
TDD · refactor · coverage · quality		superpowers:test-driven-development · code-quality
"scaffold a Next.js component"
fullstack · API · server component		backend-development · ui-design
"what's our convention for API routes?"
naming · error handling · perf rule		rules
"broad/long autonomous goal: …"
vague · stalled · multi-task		goalbuddy
"caveman: list deltas"
terse · low-token · be brief		caveman
"handoff this session"
compact · baton-pass		superpowers:handoff · handoff
"knowledge graph this codebase"
communities · viz		graphify
"schema, migration, PostgreSQL perf"
dep audit · query plan		data-engineering · database-design
Self-registration

New skill goes in ~/.claude/skills/<name>/SKILL.md — the router table auto-updates on save or next session start. No manual edit. The PostToolUse hook fires the sync; SessionStart catches anything added offline.

007
Chapter Seven · Spec-Kit + TDD

The highest-success
workflow you own.

Seven commands turn a vague idea into shipped, tested code without the usual hour-3 rework. The full visual lesson lives at ~/.claude/docs/lessons/spec-kit-tdd.html. Here's the gist.

01
Constitution
/speckit.constitution
02
Specify
/speckit.specify
03
Clarify
/speckit.clarify
04
Plan
/speckit.plan
05
Tasks
/speckit.tasks
06
Analyze
/speckit.analyze
07
Implement
/speckit.implement
The magic prompt
/speckit.implement task 1

Use the superpowers:test-driven-development skill. Red-green-refactor.
Show me the failing test before the implementation.

That second paragraph is what flips the agent from "write code that probably works" to "prove each step." Without it, you'll sometimes get implementation first, test second — which is not TDD. With it, every cycle is RED (failing test) → GREEN (minimum code) → REFACTOR (still green) → conventional commit.

008
Chapter Eight · TL;DR Runbook

How to get
the most out
of a new session.

Three habits

  1. Phrase requests with router-matching keywords.Stop pre-loading. One trigger keyword = one skill. The router does the rest.
  2. Default to spec-kit + TDD for non-trivial work.Seven commands. The magic prompt. No feature code without a spec.
  3. Feed memory when something is worth remembering."Remember that X" → I write to the right bucket. Six months of this compounds.

Weekly cadence

  1. Friday 3pm: autonomous prune fires.Monday morning, glance at memory/feedback/router-suggestions-*.md and apply what you agree with.
  2. Fire it manually anytime.Start-ScheduledTask ClaudeBrain-AutoPruneWeekly

When you add a new skill

Just drop the SKILL.md with YAML frontmatter. The router auto-registers on save. No manual CLAUDE.md edit.

# Create a new skill mkdir ~/.claude/skills/my-new-skill # Add a SKILL.md with frontmatter: --- name: my-new-skill description: What it does + when to use it ---

When something feels worse

Everything is a git commit away from a clean rollback.

# See what changed git -C ~/.claude log --oneline # Full rollback to baseline git -C ~/.claude reset --hard e0b4501 # Or just flip one plugin back to true in settings.json

When you want to see your own usage

# Skill invocations this period cat ~/.claude/telemetry/skill-invocations.jsonl \ | sort | uniq -c # Recent prompts tail -50 ~/.claude/telemetry/prompts.jsonl # Last auto-prune run tail ~/.claude/telemetry/auto-prune.log
The mental model · in one sentence

Cold-start carries identity, router, memory.
Everything else is opt-in by keyword,
observed by telemetry,
auto-pruned weekly,
self-registering when added.

— end of report —
    ╔═══════════════════════════════════════════════╗
    ║   CLAUDE-BRAIN · OPS LOG · VOL 01            ║
    ║   ────────────────────────────────────────   ║
    ║   files changed     · 47                     ║
    ║   plugins reduced   · 33 → 16                ║
    ║   skills active     · 17 → 7                 ║
    ║   cold-start saved  · ~10,000 tokens         ║
    ║   commits shipped   · 3                      ║
    ║   automations live  · 2                      ║
    ║   ────────────────────────────────────────   ║
    ║   next prune        · friday 15:00 ET        ║
    ║   rollback handle   · e0b4501                ║
    ╚═══════════════════════════════════════════════╝