Skip to content
pm-skills

Release v2.12.0. OKR Skills Launch

Released: 2026-05-03 Type: Feature release (minor) Skill count: 40 (up from 38) Key theme: OKR Skills set; full quarterly write-and-score cycle

TL;DR

You can now run a complete quarterly OKR cycle through pm-skills. The new pair:

/okr-writer    # at cycle start: draft, review, rewrite, or coach OKRs
/okr-grader    # at cycle close: score, interpret evidence, prepare next cycle

Both skills ship together so the type enum (committed | aspirational | learning | operational_health | compliance_or_safety) is consistent across the write-and-score handoff, and so users do not see a writer that points to a not-yet-existing grader.

The grader enforces the misuse failure modes that make OKRs dangerous: it refuses to retroactively change targets, refuses to retroactively shrink committed scope to claim a pass, refuses to soften committed misses with aspirational scoring, refuses to average failed guardrails into the primary objective score, and refuses to use OKR scores as individual performance ratings.

Six thread-aligned library samples (3 per skill across storevine, brainshelf, and workbench) demonstrate the full surface: aspirational sweet-spot scoring, committed-fail handling, compliance_or_safety partial-coverage, indicator-class guardrail rules, retention-thesis invalidation, and the all-three-empowerment-modes (empowered, mixed, feature-team).

No behavioral changes to any pre-existing skill (a small cross-reference cleanup in the writer now points to the grader directly, since the grader exists).

What changed

graph TD
    A[v2.11.1 tagged]
    B[2026-04-29 OKR strategy session]
    C[Approach B locked: writer + grader, defer rest]
    D[foundation-okr-writer built]
    E[Codex adversarial review on writer]
    F[measure-okr-grader built]
    G[Codex adversarial review on grader, round 1]
    H[Round 2 catches taxonomy propagation]
    I[Round 3 returns 0 findings]
    J[v2.12.0 release prep + count consistency sweep]
    L[Release-state Codex confirmation, round 1 catches reference-README defects]
    M[Round 2 catches homepage / anatomy / versioning drift]
    N[Round 3 catches versioning data accuracy, anatomy counts, date alignment]
    P[Round 4 catches release-notes audit-trail and changelog anchor]
    Q[Loop terminates: 4 rounds, MEDIUM-only audit-trail meta-findings, no HIGH/IMPORTANT, below Phase 0 termination threshold]
    O[v2.12.0 tagged]

    A --> B --> C --> D --> E --> F --> G --> H --> I --> J --> L --> M --> N --> P --> Q --> O

    style A fill:#c8e6c9
    style C fill:#fff4e1
    style I fill:#c8e6c9
    style Q fill:#c8e6c9
    style O fill:#81c784

Added

  • foundation-okr-writer with command /okr-writer. Drafts, reviews, rewrites, and coaches outcome-based OKR sets across team, department, product, product-area, and company scopes. Five entry modes are detected from user phrasing:

    • Guided (default, moderate engagement): brief diagnostic, draft, score, surface issues, ask for confirmation.
    • One-Shot via --oneshot: complete OKR set in one pass with all assumptions labeled.
    • Sustained Coach: iterative loop, one component at a time, re-scored each turn.
    • Audit Only: score and critique an existing OKR set; no new drafts unless asked.
    • Rewrite: convert flawed OKRs, feature lists, or roadmap items into outcome-shaped OKRs.

    An empowered-team diagnostic captures whether the team can change initiatives mid-cycle if KRs are not moving, and adds a conditional Disclosure section to the artifact when feature-team signals are present rather than refusing to operate.

  • measure-okr-grader with command /okr-grader. Scores completed OKRs against the canonical type enum:

    • aspirational: numeric on the 0 to 1 scale; sweet spot 0.6 to 0.7.
    • committed: pass or fail against the target; nothing below 1.0 is a pass.
    • compliance_or_safety: binary; no partial credit; no retroactive scope shrinkage.
    • operational_health: pass | fail | drift-within-tolerance against the threshold band.
    • learning: validated | invalidated | partially-validated | insufficient-evidence; no numeric score.

    Indicator class guardrail is independent of OKR type and adds a never-averaged-into-primary-score rule. Two special states prevent the most common misinterpretations: not-yet-observable for cycle-window extensions past close (90-day cohorts that finish in the next quarter) and not-yet-fully-observable for committed or compliance_or_safety KRs with partial coverage (cannot be retroactively shrunk to claim a pass on the observed subset).

  • 6 thread-aligned library samples in library/skill-output-samples/. 3 per skill (storevine, brainshelf, workbench).

    The storevine Campaigns thread now spans measure-experiment-results, foundation-okr-writer, and measure-okr-grader for a complete write-and-score arc on a single product context. Brainshelf samples demonstrate a retention-thesis invalidation: an at-scale replication study returned 1.6x vs the 3.4x observed in the original beta cohort. Workbench samples exercise the conventions the storevine and brainshelf samples do not: committed KR fail (10 of 12 onboardings, not softened to 0.83 aspirational), compliance_or_safety KR not-yet-fully-observable (1 of 3 audits completed, not retroactively scoped to claim a pass), and committed KR with guardrail indicator class.

  • docs/skills/foundation/foundation-okr-writer.md and docs/skills/measure/measure-okr-grader.md mirror pages auto-generated by scripts/generate-skill-pages.py. Both skills appear in mkdocs.yml nav.

Changed

  • foundation-okr-writer/SKILL.md cross-reference cleanups now that the grader exists: line 45 redirects scoring users to /okr-grader directly; line 184 drops the “planned for a later release” framing. No behavioral change to the writer.

  • utility-pm-skill-builder packet-format simplification (silent). The Step 5 Skill Implementation Packet was reduced from 13 to 12 items in the prior session. Nothing downstream depended on the removed item; the packet format change does not affect any shipped skill or sample. Bundled silently with the OKR launch per design.

  • README.md skill counts updated 38 → 40 (26 phase + 8 foundation + 6 utility), version badge bumped 2.11.1 → 2.12.0.

  • .claude-plugin/plugin.json and marketplace.json version bumped to 2.12.0; descriptions updated to reflect 40 skills and the new OKR Skills set.

  • library/skill-output-samples/README_SAMPLES.md updated: total samples 120 → 126, total skills 38 → 40, Browse-by-Skill table extended, all three thread tables extended with OKR rows, footer version refreshed.

Infrastructure / process

  • Phase 0 Adversarial Review Loop applied per the v2.11.0 codification. Both skills passed through Codex adversarial review before commit:

    • The writer’s review caught 1 generator-script bug (HTML attribution comment rendering as page H1 across all skill pages) and 1 nonexistent-command directive (the writer redirected scoring users to /okr-grader while the grader did not yet exist). Both resolved before the writer commit.
    • The grader’s review took 3 rounds to converge. Round 1 caught 1 HIGH finding (workbench sample taught users to violate the grader’s own compliance scoring rule by retroactively shrinking the committed scope) and 2 MEDIUM findings (/define-hypothesis nonexistent slash command in samples; OKR-type vs indicator-class taxonomy drift in SKILL.md). Round 2 caught 2 MEDIUM findings (taxonomy drift propagation to TEMPLATE.md and sample KR3 framings). Round 3 returned 0 findings. All resolved before the grader commit.

  • Em-dash sweep extension across the auto-generated docs/skills/ mirror to keep the mirror in sync with the standing no-em-dash rule applied to source SKILL.md files.

  • docs/internal/audit-ci/ consolidated into docs/internal/audit/_archived/ to flatten the audit-history tree.

Why this matters

OKRs are easy to ship as artifacts and very hard to ship as a working operating system. The literature converges on the same set of failure modes: feature-delivery KRs disguised as outcomes, fabricated baselines, missing evidence sources, naive cascading, compensation coupling, and committed misses softened by aspirational interpretation at cycle close.

The pm-skills contribution is to make those failure modes hard to ship past. The writer is a coach, not a template filler: it diagnoses empowered-team context, separates outcomes from work, applies the canonical OKR type enum, and runs a Quality Audit Rubric on every draft. The grader is an evidence interpreter, not an arithmetic engine: it scores per the type, applies indicator-class rules independently, and refuses the misuse patterns even when the user (or stakeholder) asks for them.

Both skills ship at the same time so the cross-skill hand-off is coherent at first appearance. The 6 thread samples cover the empowerment spectrum (empowered for storevine and brainshelf, mixed for workbench) and the OKR type spectrum (aspirational sweet-spot, deferred not-yet-observable, committed fail, compliance_or_safety partial coverage, indicator-class guardrail).

This is the first release where the new OKR Skills set ships with a Phase 0 Adversarial Review Loop trail in the release notes. The 3-round convergence on the grader is concrete evidence that adversarial review catches design-level pedagogical bugs (the workbench sample taught users to violate the rule the skill is supposed to enforce) that traditional review would have missed.

Validation

  • All 40 SKILL.md files pass scripts/lint-skills-frontmatter.sh
  • scripts/validate-commands.sh passes (40 / 40 commands resolve to canonical SKILL.md paths)
  • scripts/validate-agents-md.sh passes (40 skill paths matched in AGENTS.md)
  • scripts/validate-meeting-skills-family.sh passes (Meeting Skills Family contract conformance)
  • scripts/validate-version-consistency.sh passes at 2.12.0 (CHANGELOG, plugin.json, marketplace.json, README badge all aligned)
  • scripts/check-count-consistency.sh passes (40 skills, 47 commands, 9 workflows; no stale hardcoded counts in tracked files)
  • scripts/check-context-currency.sh passes (AGENTS/claude/CONTEXT.md and AGENTS/codex/CONTEXT.md both at v2.12.0)
  • scripts/generate-skill-pages.py regenerates the docs/skills mirror without diff drift
  • mkdocs build --strict passes with the new docs/reference/README.md wired into nav
  • Codex adversarial review converged on stable findings: 3 rounds for the grader (round 1: 1 HIGH + 2 MEDIUM; round 2: 2 MEDIUM taxonomy drift; round 3: 0), 1 round for the writer; 4 release-state confirmation rounds (round 1: 2 MEDIUM reference-README defects; round 2: 2 MEDIUM homepage / anatomy / versioning drift; round 3: 3 MEDIUM versioning data accuracy + anatomy counts + date alignment; round 4: 2 MEDIUM release-notes audit-trail accuracy + changelog anchor). Loop terminated after the round-4 fix per the Phase 0 codified rule “until findings stabilize below IMPORTANT severity”: MEDIUM count was 2 / 2 / 3 / 2 across rounds (no escalation, no HIGH, all findings below IMPORTANT threshold), and the round-4 findings were entirely meta-level audit-trail accuracy in the release notes themselves rather than new release-state defects. The pattern of fix-introduces-fix is exactly what the Phase 0 loop is designed to catch: each resolution can introduce new drift, and only by re-running the review do you find it.
  • Zero em-dash characters in tracked files
  • Zero references to the silently removed Quality Forecast / K/P/C/W zone classification in tracked files

Upgrade guide

If you installed pm-skills before v2.12.0, you have two options:

Option A: Stay on v2.11.1 (no action needed)

The skills you already have keep working. v2.12.0 is purely additive for the OKR work; the only modification to a pre-existing skill is two cross-reference lines in foundation-okr-writer (which itself shipped in this release).

Option B: Pull the OKR Skills set

Pull the v2.12.0 tag and the new commands appear:

/okr-writer    # at cycle start
/okr-grader    # at cycle close

Use the writer to draft or audit a quarterly OKR set. At cycle close, feed the OKR set plus final KR values to the grader for scoring, evidence interpretation, learning synthesis, and next-cycle recommendations.