task-tier-loop
Use when planning any project's work across chat, Claude Code and plain Python, to pick a tier per task and run the plan, brief, run, check, review, decide loop with .md hand-offs.
Files1
View on GitHubSKILL.mdpython_skills/llm_instructions/chat_skills/task-tier-loop/SKILL.md
GitHubTask tiers and the hand-off loop
From the user's doc "Stretching Claude Pro: A General Workflow" (a Claude chat artifact). On a flat plan the scarce thing is usage, so spend it where a mistake is expensive. The deciding question for any task: if the answer is wrong, how costly is it, and how quickly would I notice?
The user prefers Claude over local models, so tier B means a cheaper Claude model or a short focused session, not Ollama, unless they ask.
Three tiers. Pick one before opening a session; if unsure, go one tier up.
| Tier | Who | Typical tasks | Why |
|---|---|---|---|
| A | Claude, full session | planning and architecture, hard debugging, auth, payments and security code, reviewing others' output, vague requirements | a wrong answer is costly or hard to spot |
| B | cheaper or shorter session | repetitive edits, boilerplate, formatting, tests from a clear spec, docs, data clean-up, scripts with clear input and output | easy to check, cheap to redo |
| C | plain Python, no model | counting, deduping, sorting, date maths, checking a quote is in a saved page (verify.py), fetching and saving pages |
same answer every time, costs nothing |
Move a task down a tier only if you can describe the finished result in a paragraph AND check it in minutes by running it or reading a diff.
The loop (every hand-off is a .md file)
- Plan on Claude: split the goal into small tasks, save
plan.md. - Brief per task:
tasks/001.mdwith the goal in one sentence, files it may touch, what done looks like, what it must not change. - Run each task in a fresh session (or as a Python script for tier C). One task per session.
- Check with code first: tests, the script,
verify.py. Model output is not trusted until something other than the model checked it. - Claude reviews the diff only: the brief plus the change, not the whole project.
- The user decides: accept, or send back with a note in the brief file.
Research follows the same order: save the page as a .md snapshot with url and date, record each claim with an exact quote, let verify.py confirm it before anyone reads it.
Chat to Claude Code hand-off
When work moves from chat to the laptop, the hand-off is a HANDOFF.md in the repo: where we stopped, the ordered task briefs, what each check should print, and the open decisions. Claude Code starts from that file, not from chat history.
Habits
Short sessions, one task per session, brief files instead of re-explaining, show only the file or diff needed, move mechanical work to code, prefer fewer larger requests.