Skip to content

v2.33.0

Latest

Choose a tag to compare

@github-actions github-actions released this 01 Sep 17:34
· 10 commits to main since this release

The release users notice

Every fix in this release came from someone using the skills and telling us what broke.

That is a first for this project. Previous releases were largely the repository inspecting itself and finding its own problems, which is a bias worth naming: self-inspection scales and bug reports do not, so a finding with a validator behind it always looks more actionable than a report with a person behind it. These were prioritized anyway.


The three things people reported

The persona example did not match the persona template

foundation-persona shipped a template and an example, and they demonstrated different formats. If you copied the example, which is the entire purpose of an example, you got a persona that did not conform to the template the skill would then judge it against.

It is now regenerated as a filled-in instance of its own template, and verified structurally rather than by eye: all 13 template sections present, none added, no leftover authoring notes, no unfilled placeholders. The existing persona was re-rendered, not replaced, so if you were already working with it, it is still the same person.

Reported in #251. foundation-persona 2.6.1.

Prioritization assumed a big company with a long list

Three separate frustrations, each fixed the way the person reporting them suggested.

"Show the top 5 and bottom 5" is useless when you have 6 items. The highlight rule now scales below ten items instead of assuming a long backlog.

RICE Effort was expressed in person-months, which means nothing to a team of two. Effort is now capacity-weeks measured against your team's real capacity, and the skill states the conversion rather than leaving you to invent it.

Kano refused to run without a formal survey. It now recognizes two tiers of evidence, surveyed and inferred, determined by how you collected it, and refuses only when you have no research at all. How much research you have became a separate question that limits how strong a claim you may make, rather than a gate that stops you entirely.

Reported in #252. define-prioritization-framework 1.3.0.

Skills pointed you at other skills you had not installed

This is the subtle one, and the most valuable fix in the release.

Every skill was written assuming the whole library was present. So when a skill correctly decided it was the wrong tool and handed you off, it simply named the right one. If you install skills individually, which many people do, that name was a dead end: the skill correctly refused, correctly identified the alternative, and left you holding nothing you could act on.

Four skills now name the gap and inline the minimum instead of just pointing. measure-survey-analysis states the minimum viable experiment in plain language rather than naming an experiment-design skill you may not have. foundation-build-risk-review inlines the minimum for every skill it routes to, so a risk review never ends on a step you cannot take.

Worth noting how the scope shrank. The original plan named fourteen affected skills; measuring during the work cut it to four. The other ten matched a search for "mentions another skill" without actually routing to one. Saying "shared with foundation-meeting-brief" is a mention, not a handoff.

Reported in #253. Affects measure-survey-analysis, foundation-build-risk-review, discover-journey-map and define-prioritization-framework.


What is new

You can now write specs for AI features

Four new sections across three skills. They appear only when the thing you are specifying involves a model.

deliver-prd gains AI Behavior and Evaluation, pairing every behavior requirement with the evidence that actually holds it up, and giving refusal and abstention their own rows rather than leaving them implied. It also gains an Agent Execution Contract: which sources are authoritative, what must not be touched, how each requirement gets verified, and when to stop and escalate.

measure-instrumentation-spec gains Model Trace Capture, a privacy contract for prompt and completion traces. It treats minimization at two separate boundaries, before data leaves your system and again before it is durably stored, because those are genuinely different decisions with different failure modes. It also covers what happens when either one fails, what becomes of a failed trace, who may read a trace, whether that read is logged, retention, sampling, and opt-out.

develop-adr gains Model Choice inside Consequences: build versus buy versus prompt, what becomes coupled to that choice, the operating cost you are accepting, what reversing it would cost, and the observation that should reopen the decision.

Project memory has a front door

The previous release shipped project memory and then never mentioned it anywhere a user would look. Four surfaces now point at it.

Auditing the destination before adding those pointers turned out to matter, because the documentation they led to was not usable either. Four signposts to a dead end would have been worse than none.

That audit also caught a real trap. The docs twice described the memory state file as "gitignored." That is a property of this repository's ignore rules, not a property of the file. Nobody was telling you to ignore it in your own project. Anyone following those docs would have committed a file holding their initiative, decisions, and artifact paths, while the documentation assured them it was handled.

There is now a worked example you can read end to end: an interview synthesis, then a PRD that consumes it, both published. You can watch the PRD prompt get shorter because the personas and findings are already on record instead of being pasted in again.

One honest limit: this demonstrates the loop as designed, not as executed. It shows what the artifacts look like when the memory write path works. It is not proof that the write path behaves correctly at runtime.


If you are upgrading

Three skills went to a new major version: deliver-prd, develop-adr, and measure-instrumentation-spec.

They were originally typed as minor updates, reasoning that a section which only appears sometimes cannot break anyone. That reasoning was overturned before shipping, and the correction generalizes well enough to be worth stating:

Making something conditional narrows who is affected. It does not change what happens to them.

The proof was internal. Two published PRD samples describing an AI-generated summary feature carried no AI Behavior and Evaluation section, which made them non-compliant with the very skill that produced them. Both were retrofitted before this shipped.

What this means for you: if you have existing PRDs, ADRs, or instrumentation specs covering model-backed features, they are now incomplete against the current version of those skills. Documents that do not involve a model are unaffected. Nothing else here requires migration.

One known gap ships with this release. The retention section of measure-instrumentation-spec does not yet carry the evaluation-set copy question its own skill file requires. That is a content gap rather than a structural one, deferred with written reasoning to the next release.


By the numbers

  • Eight skills updated: three major, four minor, one patch
  • Sample library: grows to 213 samples across 63 skills
  • Catalog unchanged: 68 skills (30 phase + 11 foundation + 12 utility + 15 tool) and 6 sub-agents
  • No new skills, no renames, no removals

Install

Claude Code plugin (recommended)

/plugin marketplace add https://github.com/product-on-purpose/pm-skills.git
/plugin install pm-skills

From the archive

  1. Download pm-skills-v2.33.0.zip below and extract it.
  2. Run ./scripts/sync-claude.sh (or ./scripts/sync-claude.ps1 on PowerShell).
  3. Your client discovers commands in .claude/commands/ and skills in .claude/skills/.

Other editors and agents: use skills/ and commands/ directly from the extracted archive. See QUICKSTART.md for the full walkthrough.

Where to go next

Thanks to the person who filed #251, #252 and #253. All three are fixed here, at the shape each report suggested.