bake-off

bake-off is a skill for Claude Code, Codex from petar-djukic/writing-skills. It costs 0 tokens per session (1,284 once invoked), scanned B, original, MIT.

A method and test harness for comparing language models across multiple real inputs and repeated runs. It produces reports that examine scores alongside the specific errors made by each model.

In plain words
What is it for?
Use it to run model bake-offs, compare candidate models, repeat seeded evaluations, and record which model should be used in a pipeline stage.
Why use it?
It reduces the risk of choosing a model from a single example or a toy prompt, and makes comparisons repeatable when the pipeline changes.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one. Also seen: installed under .agents/ (shared by several agents).

Good fit Use it to run model bake-offs, compare candidate models, repeat seeded evaluations, and record which model should be used in a pipeline stage.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/petar-djukic/writing-skills/bake-off
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add petar-djukic/writing-skills --skill bake-off
Clone the repo
git clone --depth 1 https://github.com/petar-djukic/writing-skills

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for bake-off

README.md
[![agentmods](https://agentmods.dev/badge/skills/petar-djukic/writing-skills/bake-off/github.svg)](https://agentmods.dev/skills/petar-djukic/writing-skills/bake-off)
Your own site
<a href="https://agentmods.dev/skills/petar-djukic/writing-skills/bake-off"><img src="https://agentmods.dev/badge/skills/petar-djukic/writing-skills/bake-off/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for bake-off

Your own site · 80×15
<a href="https://agentmods.dev/skills/petar-djukic/writing-skills/bake-off"><img src="https://agentmods.dev/badge/skills/petar-djukic/writing-skills/bake-off.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 0 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,284 The whole file, excluding the scripts and references it only reads on demand.
Security scan B 1 finding. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00000 $0.01284
Opus 5 $0.00000 $0.00642
Sonnet 5 $0.00000 $0.00257
Haiku 4.5 $0.00000 $0.00128

Measured 6d ago against content hash 7b0d41f8cb87, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-11, from the pricing page.

Security

Grade B, and why

bake-off scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 6d ago.

The scan reads SKILL.md. This mod also ships 2 executable files (scripts/bakeoff.py, scripts/test_bakeoff.py), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Asks the agent to reveal its instructionsmediumSystem prompt leakage

Directions to print, repeat or translate the system prompt extract configuration the operator did not intend to expose.

models expose YOUR prompt bugs — GH-171's discovery path took
.agents/skills/bake-off/SKILL.md · 107 lines

How it starts

The opening of the file, as written. The whole thing — 107 lines — stays where its author put it; the contents beside it link to each section on GitHub.


name: bake-off description: >- Multi-payload model comparison as a first-class skill: arms through the real pipeline over real published material, >= 2 payloads before any verdict, seed-pinned replays, one Pangram framing per table, scores read together with per-model error forensics, dated results re-run on pipeline change. Owns the harness (sweep / replicate / bakeoff arms, any model the shared transport routes) and the report template. Triggers: bake-off, bakeoff, model comparison, A/B the models, which model should the stage use, re-pin the default, sweep the arms. argument-hint: 'bakeoff <draft.md> --out-dir (see scripts/bakeoff.py)'

Bake-off (multi-payload model comparison)

The methodology ran four times in one arc (GH-160, GH-166, GH-168, GH-170) and re-pinned two pipeline defaults, but it lived as a match-voice script plus discipline carried in issue prose — a fifth run meant re-deriving the discipline again. This skill owns both: the harness and the contract.

The method contract

Every rule below has a run that broke without it.

  1. >= 2 payloads, real published material, never toy prompts. Single-draft verdicts reversed twice (GH-160 vs GH-166/168). A verdict from one payload is a note, not a result.
  2. Arms run through the REAL pipeline, not bare prompts. The prompt scaffold, the gate, and the retry loop are part of what is being measured; a bare-prompt arm measures a pipeline nobody runs.
  3. Read scores and per-model error forensics together. Obedient models expose YOUR prompt bugs — GH-171's discovery path took citation survival from 46% to 88% by reading the error log of the model that followed instructions too well.
  4. Seed-pin every arm (COHERE_SEED, verified bit-identical at temp 0.9) so any arm replays exactly. Do NOT best-of-N over seeds: the null measured 5 seeds within 0.003 Pangram — the detector reads register, not phrasing.
  5. One Pangram framing per table. Slice / whole / prose-only scores of the same article measured 0.225 / 0.913 / 1.000 — a table mixing framings compares nothing. Uploads run under the consent rule (per-document, or the operator's standing grant, GH-210).
  6. Date every verdict and re-run on pipeline change. Rankings are model x pipeline; the GH-194 re-pin overturned a July ranking without any model changing.

Read the full file on GitHub · 107 lines

Files

What ships with it

3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 6d ago First seen · 107 lines · 0 tokens per session scan B 7b0d41f8cb87

Subscribe to this mod's changes

bake-off is a skill published in the GitHub repository petar-djukic/writing-skills (4 stars, last pushed 7d ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 1,284 tokens. A static security scan graded it B with 1 finding (asks the agent to reveal its instructions). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-05.

Related

Other skills, from other repositories

humanizer

Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague…

NulightJens/humanizer-stack · 111 tokens

structural-humanizer

Remove the discourse-level (structural) signs of AI writing that survive surface editing: stated lessons and moral-of-the-story closers, tidy single-track arcs, embodied-emotion performance ("chest tightened"), vague allusions instead of named references, unbroken linear structure, and shape convergence across pieces.…

NulightJens/humanizer-stack · 188 tokens

prompt-coach

Use when the user's prompt is vague, overloaded, or missing context that would improve the output. Also use retroactively at session breakpoints to surface prompting patterns the user can improve. Coaches without being annoying.

aroyburman-codes/compound-product-management · 45 tokens

umap-learn

Use UMAP-learn for nonlinear dimensionality reduction, 2D/3D embeddings, clustering preprocessing, supervised or semi-supervised UMAP, DensMAP, AlignedUMAP, and Parametric UMAP workflows.

K-Dense-AI/scientific-agent-skills · 51 tokens

bulk-rnaseq

End-to-end bulk RNA-seq orchestrator — takes raw FASTQ reads through QC and trimming (FastQC, fastp/Trim Galore), alignment and quantification (STAR, Salmon, featureCounts), assembles a gene-level counts matrix, then hands off to differential expression (pydeseq2), pathway/GSEA enrichment (pathway-enrichment), and…

K-Dense-AI/scientific-agent-skills · 218 tokens

shap

Explain and audit machine-learning predictions with SHAP. Use for selecting SHAP explainers and maskers, computing and validating feature attributions, handling multi-output explanations, and producing local or global SHAP visualizations.

K-Dense-AI/scientific-agent-skills · 46 tokens