ooda-loop — Claude Code skill: what it does, how to install
ooda-loop is a Claude Code skill that drives incident triage, latency debugging, and cloud spend cuts. The agent audits metrics, isolates root causes, and pauses for user sign-off before changes.
MIT ★ 136 updated 2026-11-05
npx skills add daniellockyer/dstack --skill ooda-loop -a claude-codeWhat ooda-loop does
The skill runs structured investigations for production web systems using the Observe, Orient, Decide, and Act loop. It starts by locking down one measurable goal, such as cutting p95 latency under 300 ms or trimming cloud spend without SLO breaches. During the Observe, Orient, and Decide phases, the agent works strictly in read-only mode. It cannot edit files, restart processes, or run load tests without authorization.
Evidence gathering relies on live signals from logs, distributed traces, MCP servers, and cloud vendor CLIs. Inventing metrics is barred by rule: if real data is unavailable, the agent halts and requests credentials. During the Decide stage, it formulates a single reversible experiment, stating the expected delta, invalidation trigger, and rollback steps. Execution happens only after explicit user approval, keeping production changes guarded.
Every cycle verifies post-action telemetry against the hypothesis. If an invalidation threshold is reached, the agent triggers an immediate rollback instead of stacking unverified fixes. A concise cycle report details observed deltas, remaining bottlenecks, and residual risks before moving forward.
What it helps with
- investigate production outages and service degradations systematically
- diagnose latency spikes and queue growth using live telemetry
- reduce cloud infrastructure spend without breaching service level objectives
- validate system changes through small reversible experiments
Example prompts for the agent
- Run ooda-loop to bring p95 latency on /checkout under 300 ms.
- Use ooda-loop to cut cloud infrastructure spend by 20% without SLO breaches.
- We have elevated 500 error rates on the API gateway. Start an ooda-loop investigation.
How to install ooda-loop
1Skills CLI (Vercel)
npx skills add daniellockyer/dstack --skill ooda-loop -a claude-code2Manually: put the files in place
~/.claude/skills/ooda-loop/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/ooda-loop/
1Skills CLI (Vercel)
npx skills add daniellockyer/dstack --skill ooda-loop -a codex2Manually: put the files in place
.agents/skills/ooda-loop/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add daniellockyer/dstack --skill ooda-loop -a cursor2Manually: put the files in place
.cursor/skills/ooda-loop/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install daniellockyer/dstack skills/ooda-loop2Manually: put the files in place
.github/skills/ooda-loop/SKILL.md1Skills CLI (Vercel)
npx skills add daniellockyer/dstack --skill ooda-loop -a gemini-cli2Platform CLI
gemini skills install https://github.com/daniellockyer/dstack --path skills/ooda-loopManually: put the files in place
.devin/skills/ooda-loop/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about ooda-loop
How to install ooda-loop in Claude Code?
Run npx skills add daniellockyer/dstack --skill ooda-loop -a claude-code in your terminal. The configuration installs to ~/.claude/skills/ooda-loop/SKILL.md.
What does the ooda-loop skill do?
It turns troubleshooting into a disciplined four-stage cycle. The agent inspects live telemetry, isolates the primary constraint, and proposes one approved, reversible action at a time.
What is required for ooda-loop to work?
The agent requires access to actual telemetry via MCP tools, cloud CLIs, monitoring APIs, or local log files. Docs or architecture reasoning alone are treated as insufficient data.
Source text
SKILL.md · MIT
--- name: ooda-loop description: "OODA loop for website performance, scalability, cost, errors, and outages: Observe, Orient, Decide (pause for approval), Act, verify, repeat. Use for ooda-loop, incidents, latency, throughput, capacity, or cloud spend." --- # OODA loop: website systems Loop **Observe → Orient → Decide → (pause) → Act → Observe** until the objective is met or the user stops. Never skip a phase. **Observe, Orient, and Decide are read-only.** No edits, deploys, restarts, scaling, or config/flag changes until Act, and only after the user approves. Anything that adds load or changes runtime state (load tests, cache warming, toggling debug logging or sampling) counts as an Act. ## Objective Before starting, pin one measurable objective the user owns (e.g. "p95 < 300 ms on /checkout", "cut infra cost 20% without SLO breach", "restore service"). If unclear, ask once. ## 1. Observe Collect real, dated, sourced data covering the objective: latency, errors, saturation, traffic, queue depth, spend, alerts, traces, recent changes. Use MCPs, vendor CLIs/APIs, observability tools, or local logs. Never invent numbers; docs and architecture reasoning alone don't count. **Data gate:** if you can't read live or exported signals, stop and ask for access by name: MCP/API/CLI for cloud systems, logs/tracing for local ones. ## 2. Orient Summarize, separating fact, inference, and unknown: - What is true now vs. the healthy/expected state - What changed (deploy, traffic, config, dependency) - Constraints (SLOs, budget, freeze windows, severity) - Blind spots - The dominant constraint: latency, errors, capacity, or cost ## 3. Decide, then pause Propose **one** action, preferring the smallest reversible experiment that tests the Orient picture: - **Action** - **Hypothesis**: which metric improves, by how much - **Invalidation**: what result means abort - **Risk / rollback** - **Tradeoffs** across perf, scale, cost, and reliability **Stop and wait for explicit approval.** If rejected, re-Orient or re-Decide. Incidents pause too, unless the user already authorized a specific mitigation. ## 4. Act Do only the approved action. Capture before/after on the Decide metrics, judge the outcome (good, mixed, bad) against the hypothesis, and note side effects. If invalidation trips, roll back and return to Observe; don't chain Acts. ## Loop - Objective met: stop; report outcome and residual risk - Otherwise: Observe again with post-Act data, then Orient and Decide (pause) - Blind spot blocked learning: close the measurement gap before any new Act ## Report (each cycle) ```markdown ## OODA cycle N: [objective] - Observe: facts (source, window), gaps - Orient: situation, dominant constraint, blind spots - Decide: action / hypothesis / invalidation / risk / tradeoffs (approved | amended | rejected) - Act: what ran, metric deltas, outcome, side effects / rollback - Next: loop | stop | need decision on … ```
© daniellockyer · MIT · daniellockyer/dstack