Narrative Read is an agent skill for turning dense source material into the kind of explanation that makes a difficult subject easy to hold in mind.
It is built for the moment when a summary is too compressed. Instead of extracting only the highlights, it reconstructs the subject for the reader: what happened or is being argued, how the pieces relate, why they matter, and what remains unresolved.
Narrative Read:
- reads the authoritative source rather than silently substituting notes or secondary material;
- chooses the narrative shape that best fits the material instead of imposing a fixed template;
- preserves important facts, numbers, complications, disagreements, dependencies, and uncertainty;
- keeps connective explanations grounded in the source;
- uses prose as the default and structure only when it improves the reader's mental map;
- adds specific safeguards for clinical material and sources that exceed one context window.
The skill remains deliberately light. It gives the model an editorial contract, not a mechanical writing formula.
Ask Codex to install the skill directly from GitHub:
$skill-installer install https://github.com/danielha/narrative-read/tree/main/skills/narrative-read
Restart Codex if the skill does not appear immediately.
This repository is also structured as a skills-only Codex plugin for local plugin testing and future directory distribution.
Invoke it explicitly:
$narrative-read Give me a full narrative recap of this transcript. Make it patient and clear enough that I can understand the whole discussion.
$narrative-read Turn this incident report into a causal narrative. Preserve the timeline, competing theories, mechanism, impact, and remaining prevention gap.
$narrative-read Explain this clinical workbench for an intelligent non-clinician. Separate established facts, interpretation, uncertainty, and questions for the treating team.
It can also be selected implicitly for requests involving comprehensive, long-form explanations of transcripts, clinical or technical material, research workbenches, planning documents, and website content.
Narrative Read is not a research engine, transcription tool, fact checker, or substitute for professional judgment. It works from the sources available to the agent. Any outside context should be clearly distinguished from case-specific claims.
For clinical material, the skill improves comprehension; it does not provide medical advice or replace the treating team.
Run the deterministic package checks:
python3 -m unittest discover -s tests -p 'test_*.py' -vThe behavioral eval set in evals/ covers intended triggers, negative controls, source fidelity, narrative coherence, uncertainty, and proportional structure. Behavioral outputs are rubric-scored rather than compared to a fixed template because the skill intentionally preserves editorial freedom.
Daniel Ha authored this skill with AI assistance and releases it under the MIT License. No third-party runtime code, private source material, or external service dependency is included.