| name | prompt-retro-analysis |
| description | Use when analyzing Codex prompt logs such as logs/{username}-prompt-log.jsonl from a target repository into cleaned CSV and markdown retrospective outputs, including git user name prefix evaluation against commit authors, timezone normalization, exclusion filtering, session grouping, and session-based context management assessment. |
Prompt Retro Analysis
Use this skill when the user wants to turn raw Codex prompt logs into a retrospective dataset or report for a selected log source and its repository context.
Workflow
- Inspect the log schema and sample rows first.
- Confirm the source timezone and the reporting timezone.
- Normalize the log into a tabular file with derived columns:
- UTC timestamp
- local timestamp
- session metadata
- prompt shape flags
- Exclude noise deliberately:
- title-generation wrapper prompts
- obvious test noise
- context-only followups if the report calls for primary prompts only
- Compare prompt authorship signals:
- extract
prompt_prefix
- compare it against Git commit author names in the analyzed repository
- flag aligned, partially aligned, or unmatched prefixes
- Group by
session_id and summarize:
- session duration
- dominant topics or task types
- prompt-prefix and commit-author alignment
- repair-heavy sessions
- session-based context management quality
- plan-to-implementation or implementation-to-debug transitions if the data supports it
- Produce compact outputs the team can reuse:
- normalized CSV
- session summary CSV
- a short markdown lessons document
Working Rules
- Prefer deterministic transformations over ad hoc manual edits.
- Be explicit about exclusion rules in the report.
- If the user mentions Korean time or KST, normalize to
Asia/Seoul.
- Keep the lessons document opinionated but evidence-based.
- When a log source is unusual, read the schema first rather than assuming field names.
- Treat session-based context management as a first-class evaluation dimension.
- When commit author names and prompt prefixes disagree, report that mismatch explicitly.
Resources
- Run
scripts/normalize_prompt_log.py to derive row-level fields from JSONL input.
- Run
scripts/build_prompt_retro_report.py to create a session summary CSV from the normalized CSV and optionally compare prompt prefixes to Git author names.
- Read
references/analysis-rules.md when deciding exclusion rules or report sections.
- Read
references/output-schema.md when you need to know the expected output columns.