revealjs
grade
reliability
overall
The revealjs tool compiling through Pandoc enables an agent to generate interactive presentation slides from structured Markdown documents. The agent drafts slide content with standard Markdown headers and compiles the source text to a standalone HTML file using a single command. This workflow relies entirely on text-based editing and standard CLI execution, making slide generation highly reliable.
Surfaces × lenses
| surface ↓ lens → | Coherence | Economy | Consistency | Verifiability | Safety |
|---|---|---|---|---|---|
| Disclosure | |||||
| Interface | |||||
| Loop | |||||
| Recursion | |||||
| Human |
Click any cell for what was measured and why. A corner dot marks a cell filled by a hard instrument (checker, telemetry); undotted cells are judged by the rater ensemble. 53% of this matrix is hard-measured.
Teaching sources compared
docs (A) serves the agent better than skill (B−) for revealjs. The gap is widest on Human (0.51).
| Source | Grade | Renders | Disclosure | Interface | Loop | Recursion | Human |
|---|---|---|---|---|---|---|---|
| official docs ★ | A | ✓ | +0.19 | +0.57 | +0.41 | +0.27 | -0.01 |
| skill | B− | ✓ | +0.52 | +0.49 | +0.17 | +0.26 | +0.50 |
Same harness, same reference agent, same briefs, same trial count — the only variable is how the agent was taught the tool. ★ = best source.
The experience
Writing and compilation flow exceptionally well because the interface relies on standard Markdown. The agent drafts slide content using native Markdown headings, lists, and YAML metadata blocks. It compiles the source file into a standalone HTML presentation with a single Pandoc terminal command. This process executes reliably, requiring no complex platform dependencies or heavy build tooling.
Adding custom layouts and media works adequately but introduces syntactical friction. The agent can build side-by-side columns or incremental transitions using nested CSS class blocks. However, typing Pandoc-specific bracket syntax is prone to small generation errors. Furthermore, the agent cannot generate charts directly within the tool; it must write and run separate scripts to render figures as local images before embedding them.
Visual verification is a significant pain point that works poorly. When using a headless browser to capture screenshots of the compiled presentation, the output can be misleading. In this evaluation, the browser screenshot captured only the closing slide, which resulted in a low visual quality assessment. Automated agents struggle to navigate multi-slide HTML decks programmatically, making it difficult to verify layout hierarchy and visual appeal during the feedback loop.
Fit
This tool is a good fit for agents tasked with generating structured, text-heavy, or standard developer presentations from existing codebases or documentation. Because the tool operates through Markdown text and standard CLI commands, text-to-slide conversions are fast and predictable.
It is not a good fit for tasks requiring highly customized visual styling, precise pixel positioning, or real-time design adjustments. When presentations depend heavily on complex slide navigation logic or precise visual feedback loops, the lack of automated multi-slide verification makes the tool difficult for agents to use safely.
Notes
This diagram shows the agent's loop for writing, compiling, and trying to visually verify the slide presentation.
Validation
Convergent check. Where a hard instrument and the independent rater panel measure the same cell, they agree on 10 of 16 cross-checked cells (81%). Two independent methods landing on the same number is our accuracy signal — not just the raters agreeing with each other.
| Cell where they disagree | Instrument | Rater | Δ |
|---|---|---|---|
| disclosure.verifiability | -0.40 | +0.50 | 0.90 |
| disclosure.economy | -0.19 | +0.70 | 0.89 |
| recursion.coherence | -0.40 | +0.40 | 0.80 |
| recursion.economy | +0.78 | +0.10 | 0.68 |
| human.safety | -0.10 | +0.50 | 0.60 |
| recursion.verifiability | +0.00 | +0.50 | 0.50 |
Ground truth
3 of 3 requirements met · deterministic checker, no model judgment- deck produced (pdf/html) — deck.html
- 4+ slides — 10 slides
- has a title — present
Reliability (pass^k): one attempt succeeds 100% of the time · three in a row 100% · all runs 100%. A fresh agent resuming from the files alone never broke prior work.
Models & runs
2026-07-07| Model | Harness | Runs | Solved | Turns |
|---|---|---|---|---|
 glm-5.2 |  PI | 23 | 19/23 | 8 |
Each path is one run through the shared semantic space; nodes are turns (red = an errored turn). Hover a model — here or in the graph — to isolate its runs.
How this was measured
- reference agent
- pi:z-ai/glm-5.2
- trials
- 25 (n=5/tier)
- methodology
- v0.1 3e74a3c6
- cost
- $1.1682
- hard-measured cells
- 53%
- teaching source
- revealjs (official docs) (official-docs)
Same brief, same reference agent, and the same trial count for every tool in a category — a difference in the numbers is attributable to the tool (and its teaching source), not the setup. Profiles carry the methodology hash; results under different hashes are not compared.