explainer-audio-dialogue-review

Files (1)

• SKILL.md

SKILL.md (first lines)

# Review one audio dialogue

Own only independent readable-dialogue acceptance. Do not write or revise
dialogue, generate VOICEVOX notation, synthesize audio, or coordinate human
interaction. In workflow use, apply [[explainer-content-workflow-base]] and read
its Artifact contract and the selected profile.

Require exact paths and digests for one run request, Content brief, Run
Manifest, Guide plan, human-readable dialogue, profile structure/evidence, and
immediate source report. Validate that the Manifest plans the Guide plan,
dialogue, review output, and their dependency relationships. Return a structured
missing/incompatible-input result rather than recovering context from prior
conversation or selecting a latest Artifact.

For standalone use, require exact `dialogue/<slug>.guide.yaml`, readable-dialogue,
and source paths and digests. Write `dialogue/<slug>.review.yaml`; do not require
workflow Artifacts or silently overwrite an existing review.

## Blind reconstruction first

After validating path identity without reading their semantic content, read only
the human-readable dialogue. Do not open the Guide plan, Content brief, source
report, structure/evidence, or prior dialogue until the blind reconstruction is
complete in the current review result draft.

Reconstruct:

- the recording's central listener question;
- the ordered causal account or reasoning chain;
- the practical judgment or procedure the listener could apply;
- the conditions and boundaries that change that judgment; and
- the point at which the promised explanation is complete.

Missing content remains missing. Do not fill gaps from general knowledge.

## Reveal and compare

Only after recording the blind reconstruction, open the exact Guide plan,
Content brief, source report, and necessary narrow source loci. Compare the
reconstruction with the Guide plan destination and every act's exit state.

Validate the Guide structure pattern against
`../explainer-audio-guide-plan/references/guide-structure-patterns.md`. Confirm
that its reasoning job matches the destination, every coverage item maps to a
named slot, the explanatory spine follows an intelligible slot progression, and
any repeated or adapted slot has a rationale. Fail a pattern that selected or
discarded coverage, weakened required depth, or hid a source qualification.

Then build a coverage ledger for every Guide plan coverage item:

- list the exact readable-dialogue turn numbers that perform it;
- verify each required explanatory facet separately;
- reject circular criteria such as "choose what is important";
- reject a term or conclusion that merely appears without its explanatory job;
- reject a representative example that does not state what it demonstrates;
- reject a decision-changing condition moved only to a source reference; and
- mark every missing or source-inconsistent facet explicitly.

Review every conversational act against the pattern its Guide plan assigned,
using `../explainer-audio-dialogue/references/conversational-act-patterns.md`.
Locate that pattern's moves by semantic function; they need not occupy one turn
each. A move performed without content fails the act, as does a closing move that
does not establish the planned exit state. An act realized with a pattern other
than the one assigned is a defect: report it and identify whether the plan or the
dialogue is the owning return point.

## Output

For workflow use, write the exact planned
`structured/dialogue-reviews/<slug>.yaml` Artifact. For standalone use, write
the sibling review path described above. Use this shape:

```yaml
schema_version: 1
review:
  manifest: {path: structured/manifest-v3.yaml, sha256: "..."}
  content_brief: {path: structured/content-brief-v2.yaml, sha256: "..."}
  source_report: {path: reports/chapter-04.md, sha256: "..."}
  guide_plan: {path: structured/guides/chapter-04.yaml, sha256: "..."}
  dialogue: {path: dialogue/chapter-04.txt, sha256: "..."}
blind_reconstruction:
  central_question: "..."
  causal_account: ["..."]
  practical_procedure: ["..."]
  decision_conditions: ["..."]
  completion_point: "..."
destination_comparison:
  destination_reconstructed: true
  act_exit_states:
    A01: pass
guide_structure:
  pattern: problem-cause-solution
  destination_fit: pass
  coverage_mapping: pass
  spine_progression: pass
  result: pass
coverage_ledger:
  G01:
    turns: [7, 8, 9, 10, 11, 12]
    facets:
      claim: pass
      mechanism: pass
      criterion: pass
      boundary: pass
    result: pass
act_pattern:
  A01:
    pattern: "<assigned catalog id>"
    moves:
      "<move name>": [7]
      "<next move>": [8, 9]
    result: pass
unresolved: []
return_to: null
result: pass
```

Use `result: pass` only when the listener destination and every planned exit
state can be reconstructed, the Guide structure passes, every coverage item and

From choplin/explainer-studio, MIT.

More from choplin/explainer-studio