no-slacking

name

no-slacking

description

Create or repair the no-slacking coding harness that lets Codex, Claude Code, and similar CLI coding agents make reliable progress across many context windows. Use when turning a product spec into an approval-gated `architecture.md` plus `task.json`, resuming an AI-built repository across multiple sessions, installing repo-level workflow files (`AGENTS.md` / `CLAUDE.md`), or running a checkpoint-or-batch automation loop with regression checks, progress logs, and immutable task status.

No-Slacking Skill

This skill packages a durable harness for multi-session coding work. It turns "build this app" into explicit project artifacts and a bounded session loop so a fresh agent can continue where the previous one stopped.

Use This Skill For

• Bootstrapping a repo from a spec so future agent sessions can continue with minimal human steering
• Repairing an AI-generated repo that lacks task tracking, handoff notes, or a repeatable startup command
• Converting a one-shot prompting workflow into a restartable checkpoint or continuous batch loop
• Installing the harness under ~/.codex/skills/ or ~/.claude/skills/

Do not use it for tiny one-off edits that clearly fit in a single session.

What The Harness Produces

When scaffolding a project, create or normalize these files in the target repo:

• architecture.md as the design baseline the user reviews first
• task.json as the source of truth for work still to do
• progress.md as the cross-session handoff log
• init.sh as an idempotent environment bootstrap
• AGENTS.md for Codex and/or CLAUDE.md for Claude Code
• automation-logs/ only when the outer loop script is used

If the repo already uses task-list.json, progress.txt, or another stable pair of names, keep those names when renaming would create more churn than value. Preserve the invariants instead of forcing the default filenames.

Operating Modes

1. Bootstrap: create harness files, decompose the spec, prepare the repo for future agent sessions
2. Resume: read the repo state, verify regressions, finish one or more tasks according to the current execution mode, and leave a clean handoff
3. Repair: harden an existing harness with vague tasks, missing instructions, or a broken init.sh
4. Automate: prefer a supervisor loop that repeatedly launches fresh coding sessions after one manual proving run

Non-Negotiable Protocol

• No implementation before the user explicitly approves architecture.md and task.json
• A bare approval reply should unlock the plan and start the default post-approval execution batch unless the user explicitly narrows scope
• One task per commit
• Regression check before new work
• Never rewrite task text during normal coding sessions
• Only flip passes after verification
• Code, task status, and handoff notes move together
• If blocked, document and stop; do not fake completion

Read references/harness-principles.md when you need the design rationale behind those rules.

How To Use This Skill

1. Classify the repo state

• If the repo has no harness, read references/initializer-workflow.md
• If the harness exists and work should continue, read references/coding-workflow.md
• If installation or CLI flags matter, read references/platform-adapters.md

2. Bootstrap or normalize the environment

• Prefer scripts/setup-harness.sh when creating the initial scaffold
• Make init.sh idempotent; later sessions must be able to rerun it safely
• Preserve the repo's current conventions when adapting an existing project

3. Draft the architecture first

• Read the spec before deciding the stack or slicing the work
• Write architecture.md with the product scope, chosen stack, major modules, data model, external integrations, delivery phases, and explicit out-of-scope items
• Capture important assumptions and unresolved risks so the user can review them before coding begins

4. Build a restartable backlog

• Tasks should usually fit in one checkpoint-sized unit of work
• Put acceptance criteria into steps
• Order by dependency, not by surface area
• Store the backlog in task.json by default and include an approval status that stays pending until the user confirms the plan
• Include a small execution policy in task.json so the repo records what a bare approval should do next
• NEVER set default_mode_after_approval to checkpoint; use continuous-batch so a bare approval keeps the agent working through the full backlog. Set default_tasks_per_run to at least the total task count (50 is a safe default) so the agent does not stop mid-backlog.
• During bootstrap or repair, you may append missing tasks if the original backlog is incomplete; once normal coding starts, task definitions become immutable and only passes should change

5. Ask for approval and stop

• After writing architecture.md and task.json, summarize the plan and ask the user whether it is reasonable
• If the user asks for changes, revise the planning artifacts and ask again
• Do not start product implementation, continuous execution, or unattended automation while approval is still pending
• If the user's next reply is a bare approval like approve, approved, looks good, go ahead, 可以, 同意, 开始, or 继续, treat that as both plan approval and authorization to begin the repo's default post-approval execution policy

6. Run the bounded coding loop

• Read architecture.md, progress, git history, and the task file
• Start the environment
• Verify one or two previously passing flows
• Select one incomplete task
• Implement and test
• Update task status and the session log
• Commit atomically, then either continue to the next task or stop cleanly if blocked
• If approval was just granted by a bare approval reply, follow the repo's configured post-approval policy, which should usually be a supervisor loop for long-running work unless the user explicitly constrained the run

7. Automate conservatively

• Start with a manual proving run before unattended runs
• When using automation-loop mode, prefer small --max-runs values first
• Review logs between batches
• If multiple runs make no progress, repair the harness before continuing
• Prefer a scriptable supervisor loop over an oversized single interactive session for truly long-running work
• Treat sub-agents as an optional acceleration technique inside one session, not as the primary portable control plane across different coding agents
• In interactive sessions, queued tasks are only backlog state until the current prompt explicitly requests checkpoint mode or continuous batch mode and the plan has already been approved

Skill-Specific Guidance

• Default artifact names are architecture.md, task.json, and progress.md
• Generate both AGENTS.md and CLAUDE.md when possible so the workflow lives in the repo, not only in the human prompt
• Keep generated project instructions concise and repo-specific; keep the reusable theory in this skill's reference files
• Use JSON for task tracking unless the repo already depends on another structured format
• Do not start unattended automation until init.sh, build/test commands, and at least one regression check are proven, and the plan approval is recorded

Installation

• Codex-style personal skill dir: ~/.codex/skills/no-slacking
• Claude Code personal skill dir: ~/.claude/skills/no-slacking
• Use scripts/install-skill.sh to symlink or copy this folder into one or both locations

Reference Map

• references/harness-principles.md: why the harness is strict and what invariants matter
• references/initializer-workflow.md: how to scaffold a repo and generate the initial backlog
• references/coding-workflow.md: checkpoint mode and continuous batch mode
• references/platform-adapters.md: current Codex / Claude Code installation paths and execution patterns