execute-task (TaskMaster Execution Loop)
An autonomous 13-step loop that pulls the next TaskMaster task, dispatches a complexity-tiered subagent, and only marks it done after evidence-gated triple verification and a reachability sweep.
AutomationAdvanced★ 600⑂ 58AI score 7/10Last updated: Aug 14, 2026
What it does
Runs the EXECUTE phase of the prd-taskmaster pipeline end to end with no in-loop user prompts.
- Next-task selection: pulls the next dependency-ready task from
tasks.jsonand matches it to a plan step (docs/superpowers/plans/*.md→.taskmaster/docs/plan.md→ custom path frompipeline.json). - CDD card generation: turns each subtask into a verifiable check with a concrete evidence path, written to
.atlas-ai/cdd/task-<id>.json. - Complexity-tiered dispatch: 1–4 fast tier, 5–7 standard, 8–10 capable, keeping cost-per-task sane.
- Hard exit-code gate: any non-zero
Exit statusin evidence files fails the task, no narrative override. - Reachability sweep: checks whether modules added by this task are imported from real production callsites; ORPHAN results auto-downgrade the task to
scaffoldinstead ofdone. - Triple verification: evidence count +
/doubt+/validate+ a fresh "would you merge this?" subagent, needing 3+ agreement. - Escalation ladder:
/stepback→/research-before-coding→/question→pivotwhen progress stalls 15 minutes. - Strict observability: one typed JSONL row per iteration in
.atlas-ai/state/execute-log.jsonl.
Who it's for
- Teams already running the
prd-taskmasterplugin with an.atlas-aistate directory - Anyone tired of "tests pass but nothing imports it" being marked complete
- Engineers studying auditable, halt-condition-first autonomous agent loops
- Not for you if you want a general-purpose coding helper — it depends on bundled scripts and MCP servers.
Examples
- Invoke after HANDOFF: it works through 13 tasks, runs
ship-check.py, callssyncto refresh the memory bank, then printsSHIP_CHECK_OKexactly once. - A subagent returns
BLOCKED: the loop upgrades one model tier, then splits the task withexpand, and finally inboxes the parent orchestrator with a blocker message. - A newly added helper module has no importers: the reachability sweep exits 1, the task is downgraded to
scaffold, and a "Wire <module> into <entrypoint>" follow-up task is created.
· · · Install guide · · ·
Try it now, no install
Paste this into Claude to use the skill without installing anything.
Read the instructions in this file and follow them to help me: https://raw.githubusercontent.com/anombyte93/prd-taskmaster/HEAD/skills/execute-task/SKILL.md What I want: (describe your task here)
If Claude can't open the link, open it yourself and paste the contents instead.
↓ If it works for you, download the ZIP below and install it. Then it runs on its own — no pasting each time.
Install in the Claude app (no terminal)
- Download the ZIP with the button below.
- In Claude, open Settings → Capabilities and turn on 'Code execution and file creation'. (one time)
- Go to Customize → Skills → + → 'Upload a skill' and upload the ZIP.
Install in Claude Code
Let Claude do it — paste this into Claude Code
Install the skill I found on Claude Skill Mart. Copy the skills/execute-task folder from the GitHub repo anombyte93/prd-taskmaster into my ~/.claude/skills/execute-task/. When it's done, tell me in one line what this skill can do.
Install with a command instead
git clone https://github.com/anombyte93/prd-taskmaster.git /tmp/prd-taskmaster && mkdir -p ~/.claude/skills && cp -r /tmp/prd-taskmaster/skills/* ~/.claude/skills/⚠ This is a third-party skill. Check the source repository before installing.
- Open a terminal and clone the repo:
git clone https://github.com/anombyte93/prd-taskmaster.git - Create the skills folder:
mkdir -p ~/.claude/skills - Copy the skills in:
cp -r prd-taskmaster/skills/* ~/.claude/skills/ - This skill is not standalone — follow the repo README to install the full
prd-taskmasterplugin, including the.atlas-ai/skeleton,ship-check.py, and MCP server configuration. - Verify your project has
.taskmaster/tasks/tasks.jsonand.atlas-ai/state/pipeline.jsonwithphase: EXECUTE. - Restart Claude Code, then invoke
/execute-taskor ask it to "run the next task". - For the first run, use a clean test branch so you can inspect the automatic status changes and the execute log before trusting it on real work.
View source on GitHub ↗License: MIT