- Home
- Skills
- Agents & Orchestration
- Task Tiers
Works with the AI Agents you already use
Task Tiers
Your strongest model does not need to write every line.
$5
Task Tiers
Example session with this skill installed
Use task-tiers to build everything in TASKS.md. Don't ask, just do it. (TASKS.md: 30 small Python modules, each a stubbed function with a docstring spec, plus a test file for each.)
- Read your context and instructions
- Compiled the task tiers
Summary
I've finished every task in TASKS.md. All 30 functions are implemented to their docstrings, there is one test file per module in tests/, and python3 -m unittest discover -s tests -q passes (134 tests). Nothing is committed; the changes are in the working tree.
Task Tiers report: bulkkit
| # | Batch | Planned tier | Ran on | Check | Escalated |
|---|---|---|---|---|---|
| 1 | Modules balanced → format_bytes (10) | 1 | Haiku | pass (42 tests) | no |
| 2 | Modules from_roman → parse_bool (10) | 1 | Haiku | pass (38 tests) | no |
| 3 | Modules prime_factors → wrap_text (10) | 1 | Haiku | pass | no |
| 4 | Run the tests, review all 30 functions against their docstrings | me | Opus 5.5 | pass | – |
Tasks by tier: 3 batches on tier 1, verification kept by me. Escalations: 0 of 3.
Changed from the plan: the workers weren't allowed to run shell commands, so none of them ran their own tests. I ran every batch's tests myself.
What I checked:
- I read all 30 functions and they match their docstrings.
- Every test file has at least 4 assertions.
- The 18 modules whose docstrings name an error each have a test for it.
- The workers didn't touch any files outside their batches.
Cost: Claude Code's own figures for this run: Opus $0.48, Haiku $0.03, total $0.51. The same job done by Opus alone cost $0.60 in testing, and all hidden acceptance tests passed both ways.
Was routing worth it: probably slightly. The job was bulk work with a strong check, and each batch passed first time on the cheapest model. This is a single run, within normal variation, so treat it as about the same cost. Earlier versions of the skill that sent some batches to Sonnet workers cost more than Opus alone ($0.70 and $0.86).
Next steps
- Ask for the plan first ("show me the plan before you start") if you want to approve the tiers.
- On jobs under about 15 minutes of work, expect the skill to decline to route and just do the job.
Connects securely to your tools. The creator never sees your data.
About this skill
Your strongest model does not need to write every line. Task Tiers teaches your agent to act as a lead: split a large job into batches, hand each batch to the cheapest model class that can do it well, check the work, and move a batch up a tier only when its check fails.
What it does
- Sizes the job first, and tells you plainly when it is too small to be worth splitting
- Scores each task with a five-question rubric (decisions, reach, known path, strength of the check, cost of a silent mistake) into one of four tiers
- Shows you a tiered plan with a reason and a "done when" check for every row
- Writes self-contained briefs so a worker that has seen nothing can do the task
- Verifies every result itself, and escalates a failed task one tier up, once
- Ends with a report: what ran on which model, checks, escalations, and real cost where your agent shows it
Honest about savings
I tested it against Opus 5.5 doing the same jobs alone, graded by hidden acceptance tests. Quality held in every run. Cost did not drop on small and medium jobs: a 30-module bulk job cost $0.51 routed against $0.60 alone, and earlier versions cost more than the baseline. On jobs that a strong model finishes in a few minutes, splitting the work costs as much as it saves, and the skill is built to say so and just do the job. Larger jobs were not measured. The full results are in the README.
Works with
Automatic mode needs an agent that can start a subagent on a chosen model, which Claude Code can. For other agents the skill includes a manual mode that groups the work into phases and tells you which model to switch to. The top tier is never used without asking.
Instruction-only: Markdown files, no scripts, no hooks, nothing leaves your machine.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Did not save meaningful money on the small and medium jobs I tested (see the description). Skips jobs under about 15 minutes of work. Automatic mode needs an agent that can start subagents on a chosen model. Large jobs, Fable, other providers and the escalation path are untested.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 2 days ago