Skill: Figure Discussion (observation → mechanism → implication → recommendation)

When to Use

USE WHEN: writing the discussion of a figure in a paper, book chapter, or task report. Every figure that appears in a deliverable MUST have an accompanying discussion block that follows this contract. Loaded by scientific-writer, book-author, narrative-framer, and the solve.task.agent workflow.

DO NOT USE for: figure generation (use generate-publication-figures), captions only (captions are a separate, terser artefact).

Core principle

A figure that appears in a deliverable without a corresponding discussion block is a wasted figure. Reviewers and readers should be able to read the figure → discussion pair and walk away knowing what the figure shows, why it shows that, what it means, and what to do about it, in that order, without having to reverse-engineer the conclusion from the plot.

This skill ports the contract from task_solve/ results.json figure_discussion[] blocks to the paper / book authoring pipeline.

The four-part contract

Every figure discussion block has exactly four parts. Each is one to three sentences. The block lives immediately after the figure in the manuscript (paper.md or chapter.md) or in the corresponding results.json field.

1. Observation — what the figure shows

State what the figure shows, with numbers. Do not describe the plot ("a downward-sloping line"); describe the result ("the iteration count drops from 28 to 16 as n_components increases from 4 to 12").

Rules:

• Always include at least one number with its unit.
• Always state the trend in scientific, not graphical, terms.
• Reference the figure number explicitly.

2. Mechanism — why it shows that

Explain the underlying physics, numerics, or algorithm. Tie back to an equation from the methods section or a NeqSim Java class.

Rules:

• Reference an equation by number (Eq. 7) or a Java class (SystemSrkEos).
• One sentence on the dominant mechanism is enough; do not hand-wave with multiple competing explanations.
• If the mechanism is unclear, say so — do not invent one.

3. Implication — what it means for the reader

State the consequence for someone using the method, the model, or the equipment. The reader should be able to act on this.

Rules:

• Phrase as engineering or scientific consequence, not as plot feature.
• Connect to the paper's claims list (Type 1) or to a primary finding (Type 2).
• Quantify the implication wherever possible.

4. Recommendation — what to do about it

A specific action. "Use SRK with classic mixing rule for systems under 50 bara" beats "results are sensitive to mixing rule".

Rules:

• Active voice, imperative mood.
• Concrete: a parameter range, a method choice, a check to perform.
• Tied to the figure — do not recommend something the figure does not support.

Block format (paper.md or chapter.md)

Insert immediately after the figure reference:

![Iteration count vs number of components for SRK and PR EOS,
showing PR converging in 30 % fewer iterations on average for
mixtures with more than 8 components.](figures/iters_vs_nc.png)

**Discussion (Figure 3).**
*Observation.* Iteration count drops from 28 to 16 as `n_components`
rises from 4 to 12 for both SRK and PR, with PR consistently 30 %
lower (Figure 3).
*Mechanism.* The PR alpha function (Eq. 7) penalises low reduced
temperatures more strongly, accelerating the early Wegstein steps in
multicomponent flashes.
*Implication.* For pipeline-grade rich gas with $n_c > 8$, PR delivers
the same density accuracy at appreciably lower CPU cost than SRK.
*Recommendation.* Use PR over SRK for gas-processing flowsheets with
rich-gas streams; reserve SRK for the cubic-only legacy modules where
parity with historical results is required.

results.json schema

The same content goes into results.json so downstream tooling (consistency checker, traceability validator) can audit the figure→claim chain:

"figure_discussion": [
  {
    "figure": "iters_vs_nc.png",
    "title": "Iteration count vs number of components",
    "observation": "Iteration count drops from 28 to 16 as n_components rises from 4 to 12; PR consistently 30 % below SRK.",
    "mechanism": "PR alpha function (Eq. 7) penalises low reduced temperatures more strongly, accelerating early Wegstein steps.",
    "implication": "PR delivers the same density accuracy at lower CPU cost than SRK for rich-gas streams with nc > 8.",
    "recommendation": "Use PR for gas-processing flowsheets with rich-gas streams; keep SRK only for legacy parity.",
    "linked_results": ["median_iterations", "convergence_rate_pr_vs_srk"],
    "linked_claim_id": "C1",
    "neqsim_classes": ["SystemPrEos", "SystemSrkEos"],
    "equation_refs": ["Eq. 7"]
  }
]

Required keys (always):

• figure — filename in figures/
• observation, mechanism, implication, recommendation — each ≤ 3 sentences

Optional keys (include when applicable):

• linked_results — keys in results.json["key_results"] (required for task reports)
• linked_claim_id — claim ID from approved_claims.json (required for Type 1 papers only; omit for book chapters and Type 2 deliverables that have no claims manifest)
• neqsim_classes — Java classes used (for traceability; recommended whenever a NeqSim class produced the figure)
• equation_refs — equation numbers in the paper (omit for book chapters that do not number equations)

Anti-patterns

The adversarial reviewer flags these. Do not produce them.

Length budget

Top journals expect 4–7 main figures. At ≤ 180 words each, the figure-discussion total is ≤ 1300 words — about half of a typical discussion section. The remaining half is the cross-figure synthesis (narrative-framer skill territory).

Validation

paperflow.py iterate --check evidence audits:

• Every figure in paper.md / chapter.md has a discussion block.
• Every block has all four parts.
• linked_claim_id exists in approved_claims.json (Type 1 papers only — skipped when no claims manifest is present, e.g. book chapters, Type 2 task reports).
• linked_results keys exist in results.json["key_results"] when results.json is present.
• neqsim_classes strings match real classes in src/main/java/ when supplied.
• No anti-pattern phrases (configurable list).

A failing check is a BLOCKER for journal-formatter packaging (Type 1). For book builds and task reports, a failing check is a WARNING — the build proceeds but the issue is logged for the next iteration.

How this skill plugs into the agents