What’s inside

One SKILL.md (the playbook) + one starter template. A preview:

Agentic Coding Platform Skills

When to use

• You work in an editor or IDE with built-in agent features and want to use them deliberately instead of clicking whichever mode is nearest
• A task spans many files — a refactor, a feature with backend plus frontend plus tests — and one-shot edits keep breaking something three files away
• You want to delegate whole tasks to the agent while keeping real control: you decide the brief, the checkpoints, and what counts as verified
• You have run two or more agents in parallel and watched them overwrite each other's work or solve the same problem twice
• Your agent's changes look finished but fail in ways a short verification pass would have caught

Workflow

1. Choose the agent mode for the task (input: task → output: mode + reason)
   Agent modes are not flavours — they are different permission levels. A question needs the lightest mode (read and answer, change nothing). A small fix needs edit mode on named files. A whole feature needs full agent mode with planning. Pick the lightest mode that can do the job and write down why: the mode decision is the first control, and it is made before the agent starts, not after it has already wandered.
2. Delegate with a task brief (input: task + mode → output: task brief)
   Write the brief the way you would brief a capable colleague: goal, context, files in scope, files out of scope, constraints, and how success will be checked. Include what "done" excludes — "do not redesign the settings page, do not change the API shape." A brief without exclusions is an invitation to remodel the whole house while fixing a tap.
3. Split multi-file work into checkpoints (input: task brief → output: checkpoint plan)
   Break the change into checkpoints that each leave the project working: first the data shape, then the logic, then the interface, then the tests. One checkpoint per coherent step, each with its own check. If a checkpoint cannot be described in one sentence, it is two checkpoints pretending to be one. The agent works checkpoint by checkpoint and stops at each — it never races ahead to "finish" while an early checkpoint is still red.
4. Run the verification loop (input: completed checkpoint → output: verification result)
   Every checkpoint ends the same way: build, tests, and a manual look at the behaviour in the running app. The loop is build → test → observe → fix → repeat, and the checkpoint only passes when all three agree. A green test suite with a broken screen is not a pass; a working screen with a failing test is not a pass either. Verification is a loop, not a single command.

Template preview

Agentic Coding Platform — Task Brief Template

One task = one brief, one checkpoint plan, one verification trail.

Mode decision

Task: {one sentence}

Mode chosen: {read-only / edit / full agent} — Reason: {why this is the lightest mode that can do the job}

Task brief

Goal: {what the finished work does}

Context: {why now, what it connects to}

In scope: {files / areas the agent may change}

Out of scope: {files / areas it must not touch}

Constraints: {API shape, style, performance, compatibility — list}

Success check: {how we will know it works, in observable terms}

Buy the full skill pack — $4.99 Pro — all 471 skills