pm-skills

Recorded example

This is a recorded run against fictional material, not a customer result or benchmark. Relay, its research and its measurements are fictional.

Codex created the fictional input, context and follow-up author decisions. The PM skill ran in Claude Code Fable and produced the review and proposed refinement. James made none of the fictional product decisions.

Both runs took place on 13 September 2026. Only Claude Code was exercised; this demonstration did not exercise the Codex runtime.

One finding from the review

A feature brief going to engineering. The review rated it ROUGH. This is one of the findings.

Draft

  • When the request completes, show “Projects archived” and remove archived projects from the active list.
  • If any projects fail, show a general error and keep the selection so the user can try again.

The supplied context says the request can succeed for some projects and fail for others.

P0: fix before implementation

The partial-failure state is contradictory and breaks the brief's own success criterion.

For this demo, Codex wrote the fictional brief and supplied the follow-up product decisions, including how to report each result.

From the proposed fix, drafted by the PM skill

The full fix includes per-result behaviour and retry rules. This is the summary rule:

After every completed request, show a summary that states three counts: newly archived, already archived, and not archived. Never show the unqualified message “Projects archived” when any project was not archived.

The refinement is proposed. It was not applied to the original brief or re-rated. It retains an additional assumption, two timeout details needing author confirmation, and unresolved P2 questions.

Read the complete sequence

  1. Product context: supplied scenario, facts and constraints.
  2. Input brief: the fictional draft going to engineering.
  3. Exact review prompt.
  4. Complete review: the unedited recorded output.
  5. Review run metadata: timing, models and hashes.
  6. Author's follow-up decisions: additional fictional choices supplied by Codex after the review.
  7. Exact refinement prompt.
  8. Complete proposed refinement: replacements and remaining questions.
  9. Refinement run metadata.

Canonical README and repeat procedure · Excerpt evidence notes

These eleven files are published verbatim from the canonical demonstration record. The excerpt notes' relative link to the internal sample copy is preserved as part of that record; the homepage example carries its approved public framing.

Runtime and source

Runtime
Claude Code 2.1.263.
Model
Alias fable, effort xhigh; primary model resolved to claude-fable-5-1. Metadata also reports a small internal Haiku call.
Skill
PM plugin 2.17.0 at the recorded product commit.

Repeat the procedure

Follow the README with the recorded product checkout and a Claude Code installation with access to Fable. Prompt downloads retain their original local paths. Rerunners replace those paths with their own checkout paths and record new prompt hashes.

The refinement was a separate invocation reconstructed from the complete saved review and explicit follow-up author decisions. For a new run, inspect the new review and make your own decisions before requesting refinement.

Output can vary even with the same inputs. This record demonstrates a workflow and an inspectable result, not measured improvement over a baseline model or prompt.