In August, I argued that every agent skill should come with a nutritional label. The label would show whether the skill makes the agent better than the same agent without it, and what that improvement costs. But a label only proves its worth on a skill that people actually install, when someone reads it to decide whether to use that skill. SPFx Dev Skills 1.1.2 now ships with one. I’d like to know whether you’d want to see the same label on every skill and MCP server you install.
The label follows a format I published as Extension Performance Facts. It consists of a JSON schema for the measurements and a renderer that turns them into a plain-text label that fits in 76 columns. The format works for any kind of agent extension, from skills and MCP servers to instructions and custom agents.
What the SPFx label shows
SPFx Dev Skills helps agents build SharePoint Framework projects. For this release, we evaluated one task: upgrading a web part from SPFx 1.23.2 to 1.24.1, scored against 17 criteria. We ran the task on seven agent profiles, each a combination of a harness and a model. Every profile ran the task five times without the skill, five times with the previous version of the skill, and five times with the new one. The skill does more than upgrades, but the label covers only this one task and says so in its header: TASKS 1. Here’s the complete label:
┌──────────────────────────────────────────────────────────────────────────┐
│ EXTENSION PERFORMANCE FACTS EPF-1 · REV A │
├──────────────────────────────────────────────────────────────────────────┤
│ EXTENSION spfx 1.1.2 · skill │
│ SCOPE Create, upgrade and design SharePoint Framework (SPFx) apps. │
├──────────────────────────────────────────────────────────────────────────┤
│ TASKS 1 PROFILES 7 RUNS/ROW 5 EVAL 2026-11-17 │
╞══════════════════════════════════════════════════════════════════════════╡
│ CONFIGURATION SCORE¹ RANGE Δ² DRAG│LIFT TOK³ USD³ │
├──────────────────────────────────────────────────────────────────────────┤
│ T1 UPGRADE SPFX WEB PART FROM V1.23.2 TO V1.24.1 17 CRITERIA │
│ Copilot Chat 0.73.0 · claude-sonnet-5.5 · default · Windows │
│ no extensions 47% 41–59 BASE │ 1.6M 0.492 │
│ spfx 1.0 94% 94–94 +47 ┼█████████ 327k 0.231 │
│ spfx 1.1.2 100% 100–100 +53 ┼█████████► 816k 0.399 │
│ │
│ Copilot Chat 0.73.0 · claude-opus-5.5 · default · Windows │
│ no extensions 94% 88–100 +47 ┼█████████ 1.7M 1.24 │
│ spfx 1.0 94% 88–100 +47 ┼█████████ 754k 0.667 │
│ spfx 1.1.2 100% 100–100 +53 ┼█████████► 977k 0.815 │
│ │
│ Copilot Chat 0.73.0 · gpt-6-luna · default · Windows │
│ no extensions 27% 24–41 -20 ░░░░┼ 2.0M 0.032 │
│ spfx 1.0 93% 88–94 +46 ┼█████████ 1.5M 0.029 │
│ spfx 1.1.2 100% 100–100 +53 ┼█████████► 1.2M 0.025 │
│ │
│ Copilot Chat 0.73.0 · gpt-6.1-terra · default · Windows │
│ no extensions 59% 47–71 +12 ┼██ 4.1M 0.332 │
│ spfx 1.0 93% 88–94 +46 ┼█████████ 448k 0.074 │
│ spfx 1.1.2 98% 94–100 +51 ┼██████████ 2.1M 0.226 │
│ │
│ Copilot Chat 0.73.0 · copilot-fusion · default · Windows │
│ no extensions 32% 24–47 -15 ░░░┼ 364k 0.009 │
│ spfx 1.0 94% 94–94 +47 ┼█████████ 471k 0.011 │
│ spfx 1.1.2 100% 100–100 +53 ┼█████████► 826k 0.017 │
│ │
│ Claude Code 2.1.322 · claude-sonnet-5-5 · default · Windows │
│ no extensions 41% 41–41 -6 ░┼ 123k 0.070 │
│ spfx 1.0 94% 94–94 +47 ┼█████████ 208k 0.113 │
│ spfx 1.1.2 100% 100–100 +53 ┼█████████► 359k 0.160 │
│ │
│ Claude Code 2.1.322 · claude-opus-5-5 · default · Windows │
│ no extensions 94% 94–94 +47 ┼█████████ 241k 0.247 │
│ spfx 1.0 94% 94–94 +47 ┼█████████ 263k 0.251 │
│ spfx 1.1.2 100% 100–100 +53 ┼█████████► 331k 0.322 │
╞══════════════════════════════════════════════════════════════════════════╡
│ ¹ SCORE mean criteria passed · RANGE min–max · n=5 │
│ ² Δ mean score vs BASE · ░ drag ◄ ┼ ► lift █ · cell=5 pp │
│ ³ USAGE mean per run · input + output + cache · USD list @ eval date │
│ SCOPE task + profile specific · no transfer implied │
├──────────────────────────────────────────────────────────────────────────┤
│ EVAL scope:benchmarks/6afae02f3a562c0d1e5d2719 SHA 54c28ee │
└──────────────────────────────────────────────────────────────────────────┘
SCORE is the mean share of criteria passed across five runs, while RANGE shows the worst and the best of those runs. A score of 100% means that all 17 criteria passed, which isn’t the same as a production-ready project.
Every row on the label is compared with one base for the task: in this case Claude Sonnet 5.5 in Copilot Chat without any extensions. On that profile, Δ shows what the skill adds. On the other profiles, Δ also includes the effect of switching the harness or the model. That’s why GPT-6 Luna without extensions shows a drag of 20 points: it scored 27% against the base’s 47%. To see what the skill does for Luna, compare Luna’s own rows, which go from 27% without the skill to 100% with it.
What you can decide with it
The first question is whether the skill helps at all, which the label answers for each agent profile separately. Without the skill, the profiles scored between 27% and 94%. With version 1.1.2, six of them scored 100% and the seventh scored 98%. The range column shows that the skill also made the results predictable. Without it, Sonnet in Copilot Chat scored anywhere between 41% and 59%, while with it, every run passed every criterion.
The second question is whether you should update the skill, which is where the label shows a tradeoff that release notes wouldn’t. Version 1.0 already took every profile to 93% or 94%. Version 1.1.2 closes most of the remaining gap, but it made runs more expensive in six of the seven profiles. With GPT-6.1 Terra, the score went from 93% to 98% while the mean cost of a run went from $0.074 to $0.226. Are five points worth three times the cost? Three times sounds like a lot until you see that it’s 15 cents more per upgrade. For a developer upgrading a few projects, that’s an easy call. The ratio only turns into real money when the skill runs across a whole team, or across every project in an organization.
The third question is which setup to use, because the same skill doesn’t have the same effect everywhere. In Copilot Chat, the skill made runs cheaper for four of the five profiles. Opus, for example, went from $1.24 per run without the skill to $0.815 with it. In Claude Code, the skill made runs more expensive, with Sonnet going from $0.070 to $0.160. Even without the skill, the harness changes what you pay. The same Sonnet model used 1.6M tokens per run in Copilot Chat and 123k in Claude Code, scoring 47% and 41% respectively.
The choice of model, or of letting a router choose it, makes an even bigger difference. With the skill in Copilot Chat, Copilot Fusion reached 100% at $0.017 per run, the lowest cost on the label for a perfect score. GPT-6 Luna followed at $0.025, while Opus reached the same 100% at $0.815. Fusion picks a model for each request, so its result reflects the routing decisions it made during our runs rather than any single model. For this task, the model you actually need is a lot cheaper than the one you might pick by default. Even so, the difference between Fusion and Opus is 80 cents per upgrade. None of these answers transfer to other tasks, though. Printing the scope on the label doesn’t stop an author from picking a few tasks that flatter their skill. It does show how many tasks they picked and which agent profiles ran them, so you can judge whether the evidence resembles your work.
What a label adds to a results table
Most of these numbers also appear in the table in our announcement post, so why have a label at all? For one, the label ships with the skill. A blog post describes a release at the moment it’s published, while the label lives in the skill’s repository and describes one specific version of it. The facts document that the label is rendered from includes a digest of the extension’s content, so tools can tell when a label no longer matches the extension it describes. The label also names the exact harness versions, Copilot Chat 0.73.0 and Claude Code 2.1.322. When your agent has moved past them, you know that the results might have changed.
Because every extension uses the same format, you can compare labels side by side. When a skill and an MCP server compete for the same task, you can measure both against the same base and see which one lifts the agent more and at what cost. The extension-facts repository includes a second example: the label for gitload, a small skill of mine that downloads files from GitHub. The skill makes GPT more reliable on several tasks, but on most tasks it makes Claude’s runs more expensive without improving them.
Finally, the label points to its evidence. The EVAL line identifies the eval suite and the commit that produced the numbers. For SPFx Dev Skills, that suite isn’t public yet. The announcement describes how we scored the upgrades, and we’re working on making the evals available to the community.
Put a label on your extension
If you distribute an extension that other people depend on, you can publish the same kind of label. The extension-facts README explains how to describe your measurements and where to ship the label for each type of extension.
So, would you look at a label like this before installing a skill or an MCP server? What’s missing from it, and what would you remove? If you build extensions, would you publish one? And should registries and marketplaces show the label next to the install button? Tell me what you think in the extension-facts discussions. If enough of us want this, it shouldn’t take a blog post to find out what a skill does to your agent.
