An open, append-only, file-based standard for capturing prompts sent to agentic coding tools. JSON Lines on disk. Mergeable across branches. Stable across versions.
Today you use Claude Code. Next year, maybe Codex. The year after, something we haven't named yet. If every tool writes its own format, every switch is a migration.
A standard makes the data ecosystem cheap. Build a dashboard once, point it at any PLF directory, and it works for any tool that writes the format.
The format is designed so two branches can never write the same file. Merge conflicts in .prompts/ are avoided by construction, not by convention.
One file per session. Append-only. Content-addressable.
Prompt text + git context + outcome. The default. What you saw above.
When redaction or .promptcellarignore matches, the prompt body drops out — but a record with {"status":"excluded"} stays so the timeline is auditable.
Any tool that produces records matching the schema in schemas/plf-1.schema.json.
Any tool that can read JSONL and tolerates unknown fields.
Encourage projects to display PLF support in their README:
Forward compatibility is non-negotiable. Readers MUST tolerate unknown fields. Writers MUST stamp version. Breaking changes require a new top-level token (plf-2) and never invalidate older records on disk.