Agent Estimation Skill

Agent Estimation

A Claude Code skill that teaches the agent to estimate its own work in **tool-call rounds** instead of fabricating human hours.

The problem

Ask Claude (or Cursor, or any other coding agent) "how long will this refactor take?" and you'll get something like "around 2-3 hours" or "roughly half a day." Confident. Specific. Made up.

The agent has no idea. It's pattern-matching on engineering blog posts written by humans, in human units, about human workflows that have nothing to do with how it executes. The training data taught it to answer in days, so it answers in days. That's the bug.

This skill replaces that frame.

What you get

When triggered, the agent:

1. Sizes the work in tool-call rounds (its native unit of work)
2. Multiplies by a risk coefficient for spec ambiguity
3. Converts to a wallclock range, never a point estimate
4. Flags external waits (CI, approvals) separately so they don't get mixed up with agent work

Example output:

Estimated: 6.5 rounds (~13-26 min agent-time). Risk: 1.3× (cross-module). External / approval waits: none.

Install

Drop the skill into your Claude Code skills directory:

git clone https://github.com/adrozdenko/agent-estimation-skill ~/.claude/skills/agent-estimation

Restart Claude Code. The skill auto-loads.

To verify:

/skills

You should see `agent-estimation` in the list.

Trigger phrases

The skill activates when you say things like:

• "estimate effort"
• "size this work"
• "how long will this take"
• "agent estimation"
• "estimate rounds"

It also runs whenever Claude is producing effort estimates inside plans, reviews, GitHub issues, or stabilization orders.

How it works

Rounds, not hours

One round = one cycle of `think → search/edit → execute/verify → fix-if-needed`.

Bucket	Rounds	What fits
trivial	1-2	Rename, typo, single-line fix
small	3-5	Multi-file mechanical change with tests