marp
grade
reliability
overall
Marp is a command-line interface tool that compiles plain Markdown files into PDF or HTML slide presentations. The tool interprets YAML frontmatter, horizontal rule slide separators, and HTML comments to control text positioning and page layout. The agent interacts with the tool entirely through file writing and command executions in the terminal.
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 marp. The gap is widest on Loop (0.23).
| Source | Grade | Renders | Disclosure | Interface | Loop | Recursion | Human |
|---|---|---|---|---|---|---|---|
| official docs ★ | A | ✓ | +0.16 | +0.61 | +0.57 | +0.23 | +0.24 |
| skill | B | ✓ | +0.38 | +0.43 | +0.34 | +0.27 | +0.08 |
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
Slide compilation works well due to its deterministic execution. The CLI executes locally under standard directories with a single, low-overhead command. The output consistently matches the raw Markdown structure, which guarantees that a newly initiated agent can resume the task from existing files without breaking prior formatting.
Layout customization works adequately but relies on precise syntax. Syntactical commands like split backgrounds or custom lead classes are declared through inline comments. If the agent makes an error in these comments, rendering faults can occur, but the tool provides structural feedback so the agent can debug slide definitions.
Asset integration works adequately but introduces workflow steps. To include visual data, the agent must write python plotting scripts or Graphviz files, execute those files to produce image assets, and then link them via standard markdown image tags. This coordinated pipeline requires the agent to complete multiple tasks to assemble a single visual slide.
Fit
Marp is a good fit for agents that need to programmatically build slide decks from structured text or raw database content. It works well when the deck contains standardized programmatic assets such as database tables, architectural diagrams, or code blocks.
It is a poor fit for tasks that require real-time visual editing, custom vector drawing, or fine-grained pixel placement. Because the final layout is compiled directly from structured source code, precise design adjustments are difficult to achieve.
Notes
The following diagram outlines the multi-asset workflow the agent uses to generate and compile a slide deck.
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 (79%). 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 | +0.60 | 1.00 |
| human.safety | -0.50 | +0.50 | 1.00 |
| human.verifiability | -0.20 | +0.80 | 1.00 |
| disclosure.economy | -0.18 | +0.70 | 0.88 |
| recursion.coherence | -0.25 | +0.50 | 0.75 |
| recursion.verifiability | +0.00 | +0.60 | 0.60 |
Ground truth
3 of 3 requirements met · deterministic checker, no model judgment- deck produced (pdf/html) — deck.pdf
- 4+ slides — 6 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 | 23/23 | 7 |
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
- $0.9867
- hard-measured cells
- 53%
- teaching source
- marp (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.