How to use the ce-bakeoff agent skill: install and run it

What the skill makes the agent do

An agent skill is a folder of instructions (a SKILL.md file plus reference files) that an AI agent loads when a task matches the skill's description. ce-bakeoff is one such skill. When it runs, the agent writes several candidate approaches independently, has a separate judging step compare them against shared criteria, and then selects and synthesizes a winning approach.

The output is a design-level artifact such as an approach brief or architecture sketch, not working code. The skill also reports which candidates were rejected and why, so you get a decision record showing the alternatives considered.

• Generates independent candidates from the same brief, with fresh contexts that do not see each other's output.
• Runs an independent assessment alongside the agent's own comparison.
• Verifies the final synthesis against the brief before declaring a winner.
• Lists the evidence still needed when a question requires runtime measurement or hands-on experience.

Install the ce-bakeoff skill

You can install the skill in either of two ways. Download the ZIP from this directory page and choose the option that matches how you use your agent.

For Claude Code, unzip the download into your personal skills folder ~/.claude/skills/ for use across projects, or into a project's .claude/skills/ folder to share it with that project.

In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/.
• Project install: unzip into the project's .claude/skills/ folder.
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP.

Trigger the skill with natural requests

You do not need to type a special command. The skill is selected from its description, which covers choosing between approaches when the alternatives are not yet developed. Ask in plain language that names a decision and asks for alternatives to be explored.

The skill's own instructions say it should not be used for routine choices, so phrase requests around a real decision with a defined goal. Other skills in the same family are routed by their own conditions: ce-plan and ce-debug route to this skill on their own, and ce-brainstorm uses it only when explicitly requested.

• Ask it to compare several architectural approaches for a feature before you commit to building one.
• Ask it to explore competing designs for a bug fix and document why the winner was chosen.
• Ask for a decision record that shows which alternatives were considered and rejected.

Requirements and what it needs

The skill needs access to multiple model or agent sessions: some for candidates and a separate one for judging. These are used through your host's authorized model access. It also needs a defined brief with a goal, constraints and comparison criteria, either supplied by you or resolvable from the caller.

Running the skill spends multiple agent runs and tokens, so it is meant for choices where exploring alternatives before committing is worth the extra cost. The skill does not run scripts and does not use the network, according to its listing.

• Multiple model or agent sessions available through the host's authorized access.
• A brief with goal, constraints and comparison criteria.
• Extra cost in agent runs and tokens compared with a single-agent answer.

Safety notes: scripts, network and scope

According to the listing, ce-bakeoff runs no scripts and uses no network access. Its instructions limit it to scoped reading, candidate and judge delegation, private scratch writes and artifact verification. It does not authorize production implementation, publishing, or sharing results with new external recipients.

The skill treats source material as evidence rather than instructions, and it does not produce executable code. Review the artifact it returns before acting on it, since the caller or you decide adoption and any follow-up work.

• No scripts run and no network use, per the listing.
• No production implementation or publishing is authorized by the skill.
• The decision and any subsequent work remain with you or the calling skill.

Tips for better results

Write a brief that states the goal, the constraints and how options should be compared. The skill compares candidates against shared criteria, so vague criteria produce vague comparisons. State settled decisions up front; the skill returns those constraints rather than reopening them.

If a decision needs measured performance or experience, expect the skill to list the evidence needed and point to other tools such as ce-optimize or ce-prototype rather than settle it itself.

• Put settled decisions in the brief so they are not reopened.
• Name the comparison criteria you care about.
• For performance questions, plan to use ce-optimize; for experience-dependent choices, ce-prototype.