Back to Skills

du-dum

pjt222
Updated Yesterday
4 views
17
2
17
View on GitHub
Metaaiapidesigndata

About

The du-dum skill implements a two-clock architecture to separate continuous, cheap observation from occasional, expensive decision-making in autonomous agents. It uses a fast clock to collect data into a digest and a slow clock that only triggers costly actions (like LLM calls) when pending work is found. This pattern is ideal for optimizing cost and performance in agent loops where most observation cycles require no action.

Quick Install

Claude Code

Recommended
Primary
npx skills add pjt222/agent-almanac -a claude-code
Plugin CommandAlternative
/plugin add https://github.com/pjt222/agent-almanac
Git CloneAlternative
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/du-dum

Copy and paste this command in Claude Code to install this skill

Documentation

Du-Dum: Batch-Then-Act Pattern

Split observe/act → 2 clocks diff freq. Fast (analysis) = cheap → writes digest. Slow (action) = reads digest → acts if pending. Digest empty → exit immediate. Zero cost idle.

Name = heartbeat: du-dum. First beat (du) observes. Second (dum) acts. Mostly only first fires.

Use When

  • Autonomous agent budget → observe often, act rare
  • Heartbeat calls LLM every tick → waste
  • Observe cheap (API read, file parse, log scan), act expensive (LLM, write, notify)
  • Decoupled fail → observe fails → last digest still valid
  • Cron agent → analysis + action = separate jobs

In

  • Required: Data sources fast clock observes (APIs, files, logs, feeds)
  • Required: Action slow clock takes when pending
  • Optional: Fast interval (default: 4h)
  • Optional: Slow interval (default: 1/day)
  • Optional: Daily cost ceiling
  • Optional: Digest format (md, JSON, YAML)

Do

Step 1: Identify 2 Clocks

Split work → observe (cheap, freq) vs act (exp, rare).

  1. List every op
  2. Classify → observe (reads → summary) or act (LLM/write/msg)
  3. Verify split: observe ≈0 marginal, act = expensive
  4. Assign freq: fast catches events, slow meets response-time
ClockCostFreqExample
Fast (analysis)Cheap: API read, parse, no LLM4-6x/dayGitHub notifs, RSS, logs
Slow (action)Exp: LLM, write1x/dayCompose reply, dashboard, alert

→ Clean split. Every op on 1 clock. Fast = no LLM. Slow = no gather.

If err: op needs both → split. Fast collects raw → digest. Slow summarizes. Digest = boundary.

Step 2: Design Digest Format

Digest = low-bandwidth msg between clocks. Compact, human-readable, parseable.

  1. Define path + format (md recommended)
  2. Header: timestamp + source meta
  3. "Pending" section → items needing action
  4. "Status" section → current state
  5. Clear empty indicator (pending: none)

Example:

# Digest — 2026-03-22T06:30:00Z

## Pending

- PR #42 needs review response (opened 2h ago, author requested feedback)
- Issue #99 has new comment from maintainer (action: reply)

## Status

- Last analyzed: 2026-03-22T06:30:00Z
- Sources checked: github-notifications, rss-feed, error-log
- Items scanned: 14
- Items pending: 2

Empty:

# Digest — 2026-03-22T06:30:00Z

## Pending

(none)

## Status

- Last analyzed: 2026-03-22T06:30:00Z
- Sources checked: github-notifications, rss-feed, error-log
- Items scanned: 8
- Items pending: 0

→ Template w/ clear pending/empty. Slow clock decides by single check.

If err: digest >50 lines → too much raw. Move details to data file, digest = summary + pointers.

Step 3: Fast Clock (Analysis)

Observation scripts on fast schedule.

  1. 1 script per source (indep failures)
  2. Each reads, extracts, appends/rewrites digest
  3. File lock / atomic write → no partial digest
  4. Log run (ts, items, errs) → separate log
  5. Never LLM / write beyond digest
# Pseudocode: analyze-notifications.sh
fetch_notifications()
filter_actionable(notifications)
format_as_digest_entries(filtered)
atomic_write(digest_path, entries)
log("analyzed {count} notifications, {pending} actionable")

Cron:

# Fast clock: analyze every 4 hours
30 */4 * * *  /path/to/analyze-notifications.sh >> /var/log/analysis.log 2>&1
0  6   * * *  /path/to/analyze-pr-status.sh     >> /var/log/analysis.log 2>&1

→ Analysis scripts update digest. Indep → 1 fails, others continue.

If err: source down → log + keep prev entries. Never clear on source fail → stale > missing.

Step 4: Slow Clock (Action)

Reads digest, decides act.

  1. Read digest (Step 0)
  2. Pending empty/"none" → exit immediate w/ log
  3. Items pending → exp op (LLM, compose)
  4. After act → clear/archive processed entries
  5. Log run (items, cost, duration)
# Pseudocode: heartbeat.sh (the slow clock)
digest = read_file(digest_path)

if digest.pending is empty:
    log("heartbeat: nothing pending, exiting")
    exit(0)

# Only reaches here if work exists
response = call_llm(digest.pending, system_prompt)
execute_actions(response)
archive_digest(digest_path)
log("heartbeat: processed {count} items, cost: {tokens} tokens")

Cron:

# Slow clock: act once per day at 7am
0 7 * * *  /path/to/heartbeat.sh >> /var/log/heartbeat.log 2>&1

→ Script exits <1s idle. Active → processes + clears.

If err: LLM fails → no clear. Items retry next cycle. Consider retry counter → no infinite retry.

Step 5: Idle Detection

Savings = idle detect. Distinguish "nothing"/"something" min overhead.

  1. Idle check = single fast op (file read + str check)
  2. Verify idle path: 0 ext calls (no API/LLM/net)
  3. Measure duration <1s
  4. Log idle differently from active
# Minimal idle check
if grep -q "^(none)$" "$DIGEST_PATH" || grep -q "pending: 0" "$DIGEST_PATH"; then
    echo "$(date -u +%FT%TZ) heartbeat: idle" >> "$LOG_PATH"
    exit 0
fi

→ Idle = 1 file read + str match. No net, no spawn.

If err: unreliable check (false pos = missed work, false neg = waste LLM) → simplify digest. Single bool field (has_pending: true/false) most reliable.

Step 6: Validate Cost Model

Calculate → confirm savings.

  1. Fast runs/day: fast_runs = 24 / fast_interval_hours
  2. Slow runs/day: typ 1
  3. Observe cost: fast_runs * cost_per_analysis_run (~$0 no LLM)
  4. Act cost: active_days_fraction * cost_per_action_run
  5. Idle cost: (1 - active_days_fraction) * cost_per_idle_check (~$0)
  6. Compare w/ original
ArchitectureDaily (active)Daily (idle)Monthly (80% idle)
Single loop (LLM every 30min)$13.74/37h$13.74/37h~$400
Du-dum (6 analyses + 1 action)$0.30$0.00~$6

→ Model shows ≥10x cheaper on idle days.

If err: no savings → (a) fast too freq, (b) fast has hidden LLM, (c) rarely idle. Du-dum wants high idle ratio. Always active → simpler polling.

Check

  • Fast/slow split clean → no LLM in fast path
  • Digest has clear empty indicator
  • Idle detect <1s, 0 ext calls
  • Fast fail → no digest corrupt (stale preserved)
  • Slow fail → no clear pending (retry next)
  • Cost model ≥10x savings idle days
  • Both clocks log runs
  • Digest bounded (archive/clear after process)

Traps

  • Digest unbounded: Append no clear → growing log. Always clear/archive after act.
  • Fast too fast: Analysis every 5min, events daily → wastes API/IO. Match freq to event rate.
  • Slow too slow: Once/day but need same-hour → too slow. Increase freq or urgent shortcut.
  • LLM in fast: Breaks cost model. Audit fast scripts → 0 LLM. Defer summary to slow.
  • Coupled fast scripts: 1 depends on another → cascade fail. Keep indep → own source, own section.
  • Silent idle log: No log → can't distinguish "running idle" vs "crashed". Always log idle.
  • Clear digest on analysis fail: Source down → no empty write. Slow would skip actual pending. Preserve last good.

  • manage-token-budget — cost framework du-dum makes practical; du-dum = pattern, budget = accounting
  • circuit-breaker-pattern — failure case (tools breaking); du-dum = normal case (nothing to do). Together: du-dum idle, circuit-breaker fail
  • observe — methodology for fast clock; du-dum structures when observes become actionable via digest
  • forage-resources — strategic explore layer; du-dum = execution rhythm
  • coordinate-reasoning — stigmergic signaling; digest = stigmergy (indirect coord via env artifact)

GitHub Repository

pjt222/agent-almanac
Path: i18n/caveman-ultra/skills/du-dum
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

Related Skills

content-collections

Meta

This skill provides a production-tested setup for Content Collections, a TypeScript-first tool that transforms Markdown/MDX files into type-safe data collections with Zod validation. Use it when building blogs, documentation sites, or content-heavy Vite + React applications to ensure type safety and automatic content validation. It covers everything from Vite plugin configuration and MDX compilation to deployment optimization and schema validation.

View skill

polymarket

Meta

This skill enables developers to build applications with the Polymarket prediction markets platform, including API integration for trading and market data. It also provides real-time data streaming via WebSocket to monitor live trades and market activity. Use it for implementing trading strategies or creating tools that process live market updates.

View skill

creating-opencode-plugins

Meta

This skill helps developers create OpenCode plugins that hook into 25+ event types like commands, files, and LSP operations. It provides the plugin structure, event API specifications, and implementation patterns for JavaScript/TypeScript modules. Use it when you need to intercept, monitor, or extend the OpenCode AI assistant's lifecycle with custom event-driven logic.

View skill

sglang

Meta

SGLang is a high-performance LLM serving framework that specializes in fast, structured generation for JSON, regex, and agentic workflows using its RadixAttention prefix caching. It delivers significantly faster inference, especially for tasks with repeated prefixes, making it ideal for complex, structured outputs and multi-turn conversations. Choose SGLang over alternatives like vLLM when you need constrained decoding or are building applications with extensive prefix sharing.

View skill