- Home
- Skills
- Agents & Orchestration
- Agent-Ops Kit: Claude Code subagents on a backlog for hours
More screenshots
Works with the AI Agents you already use
Agent-Ops Kit: Claude Code subagents on a backlog for hours
**The problem.** Subagents drift out of scope, rebuild work that already exists, and report "tests pass" when nobody checked.
$49
Agent-Ops Kit: Claude Code subagents on a backlog for hours
Example session with this skill installed
Run the mission loop on the backlog in ops/backlog.md. Write briefs to ops/briefs/ and stop on anything that needs my decision.
- Read your context and instructions
- Compiled the agent-ops kit
- Generated the document
Example run from the package (examples/bakery-webshop/, an invented small webshop), unedited:
Sub-agent report
Report: shop-14 (builder agent)
- Step zero: no hours page existed; "opening hours" only in
contact.html(table, lines 40–58). - Added
site/pages/hours.html(weekly hours, the two closed days, the short day on 31 December). - Replaced the table on
contact.htmlwith a link; added "Opening hours" to the footer partial. make build→ "built 23 pages, 0 warnings". Checkedbuild/hours.htmlat phone width in the preview: readable.- Commit
3f9c2e1"shop-14: opening-hours page with holiday closures" (3 files). - Not done: deploy. It's outward-facing; a decision card is drafted in
card.md. - Didn't work: the footer partial is also used by the old newsletter template, which now shows the link too.
Harmless, but noted. - Open question: should the hours page also show the Saturday stall at the town fair? Not in the brief, so left out.
Sidetracks
- Calls about allergens are the next most frequent question (call log, 31 in October). A short allergen page,
same pattern, about 1 hour.
Independent review
Verdict: PASS
Checked: 7 claims, 7 confirmed, 0 wrong, 0 could not check
Re-run: make build → "built 23 pages, 0 warnings"
Findings
- "No hours page existed" — confirmed: git log for site/pages/ shows no earlier hours file.
- hours.html content — confirmed: closed 25 and 26 December, open until 14:00 on 31 December, matches brief.
- Contact table replaced by link — confirmed in the diff of commit 3f9c2e1.
- Footer link — confirmed; also appears in the newsletter template, as the report says. Harmless.
- Commit is path-limited — confirmed: 3 files, all under site/.
- Nothing deployed — confirmed: the live page still shows the old contact table.
- Phone width — confirmed in preview at 375 px.
Fix list: none.
Decision card for the human
Ask: OK to deploy the new opening-hours page (shop-14) to the live webshop?
Reply: yes / no / "after Friday"
Unlocks: the holiday banner item (shop-15), which links to this page.
Built and reviewed: QA verdict PASS (qa.md), commit 3f9c2e1. Preview: build/hours.html. It replaces the
hours table on the contact page and shows 25–26 December closed and 31 December until 14:00. Cost: none.
Deploying is outward-facing, so it waits for you. If yes, I run the usual deploy script and check the live page
once. If no, the build stays local and only the banner item waits.
contact.html
PDF · document
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem. Subagents drift out of scope, rebuild work that already exists, and report "tests pass" when nobody checked.
What the skill does. Each round it:
- reads your backlog and picks 1 to 4 items;
- writes a brief file for each, containing the aim, the exact deliverable path, step zero ("check it is not already done") and the rule to stop and report at any of four bright lines;
- runs subagents in parallel;
- has a second agent with a fresh context re-check every claim against files, logs or a re-run;
- records the verdict.
Anything that touches live money, accounts or secrets, anything outward-facing, or any purchase is not acted on. It becomes a decision card you answer in one line, and the session carries on with other items.
What you get:
SKILL.md(the coordinator), with four placeholders to fill intemplates/: brief, decision card, claim-vs-artifact checklisttools/:dead_routes.py, a registry of closed lanes, andideas.py, a scored idea backlog. Both use only the Python 3.9+ standard library.docs/: change-enablement mapping, the four bright lines, a holdout-data patternexamples/: one invented run from start to finish (brief, report, verdict, decision card)
Why not just prompt it.
- The brief is a file, so the scope can be audited afterwards.
- Step zero blocks duplicate work.
- The review uses a fresh context, so the builder never grades itself.
- There are at most two fix-and-review rounds.
Known limits.
- It tells agents when to stop. It does not technically block them.
- You fill in four placeholders before the first run.
- There are no services, no accounts and no telemetry.
What the screenshots show:
- The brief template: every subagent gets a brief with the exact deliverable path, a step-zero "not already done" check, and stop and report at four bright lines.
- A reviewer verdict from the worked example: a fresh-context reviewer re-checks each claim before anything counts as done.
- A decision card from the worked example: publishing, money, accounts or a purchase wait for your one-line answer.
The examples are invented and are labelled as invented.
Pairs with Verify Agent Work on Agensi, a dedicated reviewer skill. Independent work; not affiliated with or endorsed by Anthropic.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 5 days ago
- Software
Git
Python 3.9+