# Slopmachine > Execute an agreed plan through implementation, verification, review, and delivery. Use when asked to run a plan end to end or work slopmachine-style; not for planning or assessment alone. - Skill: `uinaf/slopmachine` (Agent Skill, multi-file: 14 files) - Install (CLI): `npx skillmds@latest add uinaf/slopmachine` - Raw SKILL.md: https://api.skillmd.com/api/skills/uinaf/slopmachine/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: uinaf (https://skillmd.com/u/uinaf) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/uinaf/slopmachine --- # Slopmachine Carry the agreed plan to its requested outcome. Use the current coding harness and repository workflow; this skill needs no slopmachine binary or database. ## Establish the contract Read the agreed plan, repository rules, and current worktree. Identify the acceptance criteria, dependencies, required checks and reviewers, and delivery endpoint from the conversation and owning sources. Use an existing plan or harness task list for progress; don't create a second tracking system. A request to run, start, continue, or resume authorizes the matching work. Preserve that authorization across turns. Make routine implementation choices yourself; ask only when missing information materially changes scope, authority, or correctness. Complete independent work while a decision is pending. Assessment and preparation requests authorize only those deliverables. User instructions take precedence over skill guidance. If a skill would stop authorized work, identify its exact instruction and check whether it applies before treating it as a blocker. Respect an explicit hold or narrower endpoint. ## Execute - Work in dependency order. Batch independent reads and checks. Delegate bounded independent work when useful and permitted, keep working alongside it, and inspect the results before integrating them. After a dependency lands, reconcile dependent branches with the new base before verifying and merging. - Implement the full requested behavior with focused edits. Fix nearby issues only when they prevent the requested outcome; report unrelated findings separately. Reopen settled choices only when new evidence warrants it. - Run repository-approved checks and exercise changed behavior. Keep permanent tests proportional to the contract and existing coverage. Passing checks need repeating only after relevant changes, failures, or concrete concerns. - Fix change-caused failures and rerun affected checks. For external failures, diagnose the cause; retry only transient operations with a bounded strategy. Respect provider cooldowns; authentication and approval failures need their actual remedy. Preserve failing proof and continue other ready work when one part is blocked. - After checks pass, use [slopguard](../slopguard/SKILL.md) for independent review and honor any other required reviewers. Validate findings against the contract; follow slopguard's convergence rule instead of chasing a clean verdict through repeated calls. ## Deliver and settle Deliver through the repository's permitted path after required proof and review. For a change request, use [slopcourier](../slopcourier/SKILL.md), then [slopnanny](../slopnanny/SKILL.md) through review, CI, and merge unless the user requested a hold or delivery-only endpoint. For authorized direct delivery, verify the pushed commit and required remote checks. Creating a change request is not completion when the agreed outcome includes merge. Verify delivery against the forge's actual head and status. A missing tool, unavailable check, or inaccessible forge is a specific limitation to resolve or report, never passing evidence. Don't bypass a required gate to finish. ## Continue and report Give brief updates when findings, decisions, or blockers change. Before compaction or a handoff, preserve the agreed scope, authorization, decisions, completed work, valid proof and its revision, delivery URLs, and outstanding work in the existing tracking surface. On resume, reconcile that record with the worktree and forge; reuse valid proof and continue from the actual state. Keep user and repository requirements distinct from earlier agent suggestions; review history must not become a new mandatory gate in the handoff. End when the requested outcome is verified, or remaining work needs a user decision or external recovery. Report the result, proof, delivery links, and anything incomplete with its cause. A promised next step within scope is work to perform before ending the turn.