talont-org/plan-visualization

Plan Visualization Skill

Reads the finalized experiment plan, selects 2–4 vis-lens skills via three-tier logic, runs them in parallel, resolves conflicts across their yaml:figure-spec outputs, and synthesizes a complete visualization plan.

When to Use

• As the plan_visualization step of the research recipe, after review_design GO and before create_worktree

Arguments

/autoskillit:plan-visualization {source_dir} {experiment_plan_path} {scope_report_path}

• {source_dir} — Absolute path to the source repo (the CWD before worktree creation)
• {experiment_plan_path} — Absolute path to the finalized experiment plan markdown
• {scope_report_path} — Absolute path to the scope report (may be empty string if absent)

Critical Constraints

NEVER:

• Select fewer than 2 or more than 4 lenses
• Skip vis-lens-always-on (it is always Tier A)
• Run vis-lens calls across multiple assistant messages — all selected lens calls must appear in a SINGLE assistant message to execute in parallel
• Write outputs outside {{AUTOSKILLIT_TEMP}}/plan-visualization/
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use model: "sonnet" when spawning all subagents via the Task tool
• Write a vis-lens context file for each selected lens before invoking it
• Log every conflict resolution decision in the Conflict Resolution Log table
• Emit visualization_plan_path and report_plan_path tokens as your final output

Workflow

Step 0 — Parse Arguments

Extract source_dir, experiment_plan_path, and scope_report_path from arguments. Read the experiment plan at experiment_plan_path.

Extract the following fields (use sensible defaults if absent):

• experiment_type — string (e.g., "benchmark", "ablation", "correlation")
• training_curves — boolean (default: false)
• num_DVs — integer count of dependent variables (default: 1)
• comparative — boolean (true if multiple conditions compared head-to-head)
• DV_types — list of DV type strings (e.g., ["accuracy", "temporal", "latency"])
• num_conditions — integer count of experimental conditions (default: 1)
• target_domain — string (e.g., "nlp", "cv", "rl", "general")

Step 1 — Three-Tier Lens Selection

Build selected_lenses (list of 2–4 vis-lens skill slugs):

Tier A (always selected, mandatory):

• vis-lens-always-on

Tier B (select 1–2 based on experiment_type and override rules):

Override rules (checked first, in priority order):

1. If training_curves == true → include vis-lens-temporal
2. If num_DVs >= 6 AND comparative == true → include vis-lens-multi-compare
3. If DV_types contains "temporal" → include vis-lens-temporal (if not already)
4. If num_conditions >= 8 → include vis-lens-multi-compare (if not already)

Experiment-type table (use when no override fires or to fill second Tier-B slot): | experiment_type | Primary lens | Secondary lens (optional) | |---|---|---| | benchmark | vis-lens-chart-select | vis-lens-uncertainty | | ablation | vis-lens-multi-compare | vis-lens-chart-select | | correlation | vis-lens-chart-select | vis-lens-figure-table | | regression | vis-lens-temporal | vis-lens-uncertainty | | classification | vis-lens-chart-select | vis-lens-uncertainty | | (default) | vis-lens-chart-select | — |

Cap Tier B at 2 lenses total (overrides count toward this cap).

Tier C (0–1 based on methodology tradition detection):

Tier C selects vis-lens-methodology-norms when the experiment plan's research methodology is identifiable from the 12 bundled methodology traditions.

Stage 1 — Deterministic keyword match:

1. Load all methodology traditions from recipes/methodology-traditions/*.yaml
2. For each tradition, count how many of its detection_keywords appear in the experiment plan text (case-insensitive, word-boundary matching)
3. Build candidate_set = traditions with ≥ 2 keyword matches
4. Branch on len(candidate_set):

| Candidates | Action | precedence_trace | |---|---|---| | 0 | Skip Tier C entirely | stage1_no_match_fallback | | 1 | Use that tradition as primary_tradition | stage1_single_match | | ≥ 2 | Proceed to Stage 2 | — |

Stage 2 — LLM arbitration (only when ≥ 2 candidates):

1. If any registered UnionRuleDef covers the candidate set, apply it: select resolved_tradition, record rule name in applied_union_rules, set precedence_trace = "stage2_tiebreak_by_rule_{rule_name}"
2. Otherwise, select among candidates by analyzing the plan's primary research question and methodology at temperature 0. Prefer the tradition whose mandatory figures are most relevant to the stated research design. Set precedence_trace = "stage2_tiebreak_by_methodology_fit"

Emit routing triple (include as a fenced yaml block in the vis-lens context file):

primary_tradition: <tradition_slug>
applied_union_rules: [<rule_name>, ...]
precedence_trace: "<trace_value>"
candidate_set: [<tradition_slugs>]

When primary_tradition is set, add vis-lens-methodology-norms to selected_lenses and write the tradition_slug and routing triple into its context file (Step 2).

Only add Tier C lens if it is not already in Tier A or Tier B.

Enforcement: Total must be 2–4. If total < 2, add vis-lens-chart-select. If total

4, drop the last Tier C lens, then last Tier B secondary.

Step 2 — Write Vis-Lens Context Files

For each lens in selected_lenses, write a context file:

Path: {{AUTOSKILLIT_TEMP}}/plan-visualization/vis_ctx_{slug}_{YYYY-MM-DD_HHMMSS}.md

Template for each context file:

# Vis-Lens Context: {slug}

## Experiment Summary
{1–3 sentence description of the experiment from the plan}

## Data Shape
- Dependent Variables ({num_DVs} total): {DV names and types}
- Independent Variables: {IV names, levels, and ranges}
- Conditions: {num_conditions} conditions
- Replication: {n_seeds or n_trials if available}

## DV Specification
{For each DV: name, type (continuous/discrete/temporal), unit, expected range}

## IV Specification
{For each IV: name, type, levels (for categorical) or range (for continuous)}

## Comparison Structure
- Comparative: {true/false}
- Head-to-head pairs: {list if applicable}
- Factorial interactions: {list if applicable}

## Expected Data Outputs
{List the files or data structures the experiment will produce, from the plan's
data_manifest or results/ section if available}

When the context file is for vis-lens-methodology-norms, append the following section to the template above:

## Methodology Tradition
tradition_slug: {primary_tradition from Tier-C routing triple}
routing_triple:
  primary_tradition: {slug}
  applied_union_rules: [{rules}]
  precedence_trace: {trace}
  candidate_set: [{candidates}]

Step 3 — Run Vis-Lens Skills in Parallel

In a single assistant message, invoke all selected_lenses as slash commands:

/autoskillit:vis-lens-{slug1} {source_dir} {vis_ctx_path_for_slug1}
/autoskillit:vis-lens-{slug2} {source_dir} {vis_ctx_path_for_slug2}
...

Wait for all lens outputs. Read each lens's output file (the yaml:figure-spec blocks within each lens's output markdown).

Empty plan handling: If vis-lens-always-on returns SKIP: no_figures_needed, record zero figures and proceed to Step 4 with an empty figure list.

Step 4 — Resolve Conflicts

For each figure-spec block where two lenses disagree on chart type, color encoding, or layout, apply the conflict resolution hierarchy:

accessibility > anti-pattern > methodology-norms > chart-select

Resolution rules:

• accessibility (from vis-lens-always-on or vis-lens-color-access) wins over all
• anti-pattern findings (from vis-lens-antipattern or always-on pass 1) override chart-select and methodology-norms recommendations
• methodology-norms (from vis-lens-methodology-norms) overrides chart-select
• chart-select (from vis-lens-chart-select) is the lowest priority

Every resolution must be logged as a row in the Conflict Resolution Log table.

Step 5 — Write visualization-plan.md

Path: {{AUTOSKILLIT_TEMP}}/plan-visualization/visualization-plan.md

Content structure:

# Visualization Plan

## Figure Inventory

| Fig ID | Title | Lens Source | Chart Type | Data Source | Priority |
|--------|-------|-------------|------------|-------------|----------|
| fig-1  | ...   | ...         | ...        | ...         | P0/P1/P2 |

## Figure Specifications

{For each figure: paste the yaml:figure-spec block from the winning lens}

## Code Allocation Hints

{For each figure: note which module/file the plotting script should live in,
e.g., `research/{slug}/scripts/fig1_training_curves.py`}

## Conflict Resolution Log

| Fig ID | Dimension | Lens A | Lens A Rec | Lens B | Lens B Rec | Winner | Reason |
|--------|-----------|--------|------------|--------|------------|--------|--------|

Step 6 — Write report-plan.md

Path: {{AUTOSKILLIT_TEMP}}/plan-visualization/report-plan.md

Content structure:

# Report Plan

## Section Outline

| Report Section | Figure IDs | Notes |
|---|---|---|
| Executive Summary | — | no figures in summary |
| Results | fig-1, fig-2 | ... |
| Analysis | fig-3 | ... |
| Appendix | all | full captions |

Step 6.5 — Write visualization-plan-trace.md

After completing Tier-C routing, write a trace file capturing routing decisions:

Path: {{AUTOSKILLIT_TEMP}}/plan-visualization/visualization-plan-trace.md

Content structure:

# Visualization Plan Trace

## Tier-C Routing Decision

- **tier_c_lens**: {selected_lens_name or null}
- **primary_tradition**: {methodology_tradition_slug or null}
- **disambiguation_rule_applied**: {rule_name or null}
- **applied_union_rules**: [{list of union rules applied, or empty if none}]
- **precedence_trace**: [{chain of precedence resolution, or null}]

Fill in the fields from the Tier-C routing performed in Step 1:

• tier_c_lens: the lens selected by Tier-C (0-or-1), or null if target_domain was general/others
• primary_tradition: the target_domain value that determined Tier-C selection (e.g., nlp, cv, rl, general)
• disambiguation_rule_applied: null for now (disambiguation rules from #844 not yet implemented)
• applied_union_rules: empty list [] for now
• precedence_trace: null for now

Step 7 — Emit Structured Tokens

visualization_plan_path = {absolute_path_to_visualization-plan.md}
report_plan_path = {absolute_path_to_report-plan.md}
disambiguation_rule_applied = {disambiguation_rule_applied or null}
tier_c_lens = {tier_c_lens or null}
methodology_tradition = {primary_tradition or null}
visualization_plan_trace_path = {absolute_path_to_visualization-plan-trace.md}