Agent skills¶
Draft.js exporter ships Agent Skills that help users get better results with agentic coding. They’re published in multiple ways so you can reuse them easily with a wide range of tools.
Direct link to the skill¶
Nice and simple - it’s here: https://wagtail.github.io/draftjs_exporter/.well-known/agent-skills/draftjs-exporter/SKILL.md.
Open the URL to read the skill, point an agent at it, or download it and place it in .agents/skills/draftjs-exporter/SKILL.md in your project so tooling that reads local skills picks it up.
Well Known Discovery¶
Machine-readable index of all skills: /.well-known/agent-skills/. This is per the Well Known Discovery RFC. Agents that support the format can fetch it to discover the skills.
AI catalog¶
Machine-readable index that also covers other options than skills: /.well-known/ai-catalog.json. This is per the AI Catalog specification.
Library Skills¶
Because the skill is bundled inside the exported package, it lands in your environment at site-packages/draftjs_exporter/.agents/skills/draftjs-exporter/SKILL.md when you install the exporter.
Library Skills scans installed packages for these bundled skills and installs them into your project. It symlinks to the installed package, so the skill updates automatically whenever you upgrade the exporter.
Does the skill help¶
Yes. We evaluate the skill against the published docs on a set of prompts that ask a model to write complete, runnable draftjs_exporter code, then run every answer against the real library. The skill improves the pass rate on every model we tested, and helps weaker models the most – with the skill, all Claude models land within 5 points of each other.
These figures are a snapshot from October 2026, against draftjs_exporter 7.1.0. They are aggregates over several repeats of 35 prompts, not a guarantee – treat them as an indication of the uplift, not a ranking.
Pass rate on Claude models, skill versus no reference at all:
| Model | No docs (baseline) | Docs only | Skill only | Docs + skill |
|---|---|---|---|---|
| Haiku 5.5 | 33.5% | 66.1% | 89.4% | 90.6% |
| Sonnet 5.5 | 49.0% | 66.1% | 91.8% | 94.3% |
| Opus 5.5 | 77.1% | 91.8% | 93.9% | 95.9% |
| Fable 5.1 | 58.1% | 79.0% | 94.3% | 93.3% |
Five open-weight models (DeepSeek V4.1 Flash, DeepSeek V4 Flash Speed, GLM 5.3, GLM 5.3 Flash, Mimo V2.6 Pro) showed the same pattern, with larger uplifts over their weaker baselines:
| Model | No docs (baseline) | Docs only | Skill only | Docs + skill |
|---|---|---|---|---|
| DeepSeek V4.1 Flash | 24.8% | 78.1% | 87.6% | 89.5% |
| DeepSeek V4 Flash Speed | 12.4% | 74.3% | 74.3% | 84.8% |
| GLM 5.3 | 8.6% | 76.2% | 67.6% | 85.7% |
| GLM 5.3 Flash | 26.7% | 82.9% | 74.3% | 91.4% |
| Mimo V2.6 Pro | 10.5% | 82.9% | 88.6% | 93.3% |
Skill improvements¶
How much the skill changes things, compared with the model answering from memory (no docs, no skill). Each cell shows the baseline value, the value with the skill, and the change in parentheses. Pass rate is in points, duration in seconds per answer, cost in US cents per answer, energy in watt-hours per answer.
| Model | Pass rate | Duration | Cost | Energy (Wh/answer) |
|---|---|---|---|---|
| Haiku 5.5 | +55.9 pts | 13.2s → 9.4s (−29%) | 0.16¢ → 0.27¢ (+69%) | – |
| Sonnet 5.5 | +42.9 pts | 13.1s → 8.9s (−32%) | 1.6¢ → 3.2¢ (+100%) | – |
| Opus 5.5 | +16.7 pts | 18.9s → 9.3s (−51%) | 3.8¢ → 4.8¢ (+26%) | – |
| Fable 5.1 | +36.2 pts | 44.0s → 13.5s (−69%) | 25.1¢ → 19.8¢ (−21%) | – |
| DeepSeek V4.1 Flash | +62.9 pts | 66.5s → 22.3s (−66%) | 0.18¢ → 0.05¢ (−71%) | 0.22 Wh → 0.21 Wh (−3%) |
| GLM 5.3 | +59.0 pts | 180.3s → 64.5s (−64%) | 3.0¢ → 1.1¢ (−62%) | 7.18 Wh → 3.77 Wh (−47%) |
| GLM 5.3 Flash | +47.6 pts | 130.1s → 34.9s (−73%) | 0.22¢ → 0.07¢ (−68%) | 0.53 Wh → 0.34 Wh (−37%) |
| Mimo V2.6 Pro | +78.1 pts | 98.4s → 42.7s (−57%) | 0.67¢ → 0.31¢ (−53%) | 2.79 Wh → 2.52 Wh (−10%) |
The biggest pass-rate gains are on the weakest baselines: Haiku and the open-weight models gain 48 to 78 points, while Opus, already strong on its own, gains the least.
The open-weight costs are computed from each model's list rates on models.dev applied to the tokens the eval recorded. Without the skill these models spend most of their budget on reasoning tokens, so the skill usually costs less, not more.
The skill is usually faster even though it costs a few extra turns to load, because the model stops guessing and spends fewer tokens. On Fable and most of the open-weight models the skill is also cheaper, because the no-skill answers spend far more on reasoning or run much longer.
Energy is measured per arm over three repeats of the 35 cases (105 answers), against the provider's own accounting, for the four open-weight models only. It tracks duration more closely than cost: the longest-running, most reasoning-heavy models use the most, and GLM 5.3 more than halves its energy with the skill. Energy is not reported for the Claude models.
What the results suggest in practice:
- Use the skill if you use an agent to write
draftjs_exportercode. Without any reference, models lean on dated knowledge: they import the olddraftjs_exporter_markdownpackage, invent APIs, or claim the library has no Markdown importer. The skill prevents most of that. - The skill is usually enough. Adding the docs on top changes the Claude pass rate by −1 to +2.4 points, within one SD. On the open-weight models the docs help a little more, especially for GLM (+17 to +18 points).
- The skill is the cheapest and fastest option. It is the fastest arm on every model (9 to 14 seconds per answer on Claude) and gives the shortest answers, faster than the baseline even though it takes a few extra turns to load.
- A small model with the skill can beat a large model without it. Haiku with the skill scores higher than Opus without any reference, at roughly 1/14 of the cost per answer.
A marginal caveat: when both the docs and the skill are available, smaller models sometimes skip the skill and search the docs instead, which measures slightly lower. The skill alone is the simpler setup.
The method behind these numbers – the arms we compare, how answers are graded, and why we repeat runs – is in the skill and docs evals guide for contributors.