3 production-grade SKILL.md files that fix the three problems that kill long agent runs: prompt drift, approval queue backlog, and context window crashes.
Each skill file is a standalone context engineering protocol. Drop it in, use it immediately.
Sliding window anchor pattern that pins your original task goal at high attention weight across 50+ tool calls. Agents stop drifting. Work stays on target.
Structured review protocol with a 3-question decision tree. Clears borderline Water Fountain escalations in under 2 minutes. Includes Telegram notification formatter.
Pre-limit compaction protocol triggered at 68% context usage. Preserves decision trail, active constraints, and task state in <800 tokens. Never lose a long run to a context crash.
Every skill ships with drop-in Python and prompt snippets — built from the actual CLAW-ROOM.OS daemon implementation.
# Anchor injection — prepend to every CLAW task prompt anchor_block = f""" TASK ANCHOR [immutable — do not summarise away]: Goal: {one_sentence_terminal_output} Constraints: {hard_limits} Success signal: {exact_completion_condition} """ # Checkpoint every 10 tool calls def emit_checkpoint(call_n: int, anchor: str, state: str) -> str: return f""" ## CHECKPOINT [call #{call_n}] Anchor goal: {anchor} Current state: {state} Drift check: does output still serve the anchor? yes/no + why Next action: {next_step} """ # Bottom-of-prompt anchor reminder (for prompts > 2000 tokens) reminder = f"Recall: your goal is {anchor_goal}. Output must satisfy: {success_signal}."
async def compact_memory( context: str, anchor: str, client: anthropic.Anthropic, ) -> str: # Trigger at 68% — before the wall, not after system = ( "Compress context into MEMORY COMPACT format. " "Remove reasoning, keep decisions, constraints, " "irreversible work, and critical values. <800 tokens." ) response = client.messages.create( model="claude-haiku-4-5-20251001", max_tokens=900, system=system, messages=[{"role": "user", "content": f"GOAL:\n{anchor}\n\nCONTEXT:\n{context[-8000:]}"}], ) return response.content[0].text COMPACT_TRIGGER_RATIO = 0.68 # env-tunable
# Group the queue by risk before reviewing def triage_queue(pending: list[dict]) -> dict: return { "near_pass": [t for t in pending if 0.75 <= t["confidence"] < 0.85], "borderline": [t for t in pending if 0.65 <= t["confidence"] < 0.75], "blockers": [t for t in pending if t["confidence"] < 0.65], } # Structured correction format (parseable by downstream agents) correction = """ CORRECTION [task_id: {task_id}] KEEP: {what_was_right} CHANGE: {specific_change} CONSTRAINT: {violated_constraint} """
cp *.skill.md .claude/skills/You have agents running overnight and waking up to drifted outputs and crashed contexts. These skills stop that.
You're running complex multi-step Claude Code sessions and hitting the limits of what default prompting can do at scale.
You've got the free 2-agent kit running and want to extend it with production-grade context engineering patterns.
Casual Claude users who don't run multi-step agent workflows. If you're not hitting context drift, queue backlog, or context limit crashes, you don't need this yet.
cp *.skill.md .claude/skills/.
.skill.md files into your project's .claude/skills/ directory. Claude Code picks them up on the next session start — no config, no framework, no API calls. The Python snippets in each file are optional extras for daemon operators who want to automate the patterns.
dropshipit369@gmail.com with subject line [Skills Pack] refund and it's done. No questions, no hoops.
3 skill files. One payment. No subscription.