Score And Retro

Score and Retro

When to use

End-of-cycle scoring plus keep/kill/revise — checking whether hitting numbers served the objective, not just grading KRs.

What this skill does not do

• Does not set next-cycle targets — route to /okr:set-targets.
• Does not invent actuals — INPUT NEEDED when missing.
• Does not default objective call to KR average unless profile explicitly requires it.

Preconditions

| Input | If missing | |---|---| | Baseline, target, actual per KR | Score INPUT NEEDED — do not guess | | Org + OKR practice profiles | Use default scoring doctrine; tag [PROVISIONAL — configure profile] | | Commit/aspirational labels per KR | Ask; do not infer from score |

Provisional mode

[PROVISIONAL — configure profile for org-calibrated scoring] when profiles missing; default formula applies.

Trust spine

• Confidence bands (governance-tracking):
• High: Every KR scored with shown arithmetic; objective calls with divergence checks; sandbagging named if present.
• Medium: Some INPUT NEEDED; provisional profile scoring.
• Low: Actuals missing for majority of KRs — halt retro conclusions.
• Failure modes:
• Strategic advice vs. support: Keep/kill/revise are recommendations; calibration session owns final calls.
• Client confidentiality: Scores may be sensitive — CONFIDENTIAL header.
• Accountability gap: KR/objective divergences flagged, not averaged away.
• Analytical Rigor: N/A — governance shape.
• Incentive Gaming: Guards grade inflation and sandbagging normalization — 1.0 on aspirational or consistent ~1.0 across teams named as pattern.
• Escalation triggers: Widespread sandbagging pattern — flag for calibration, not praise.

Scoring doctrine (inline)

Apply formula from practice profile. If none set, use defaults below.

Default formula — increase-is-better KRs

score = (actual − baseline) ÷ (target − baseline)

Cap at 1.0 unless profile allows overachievement. Below baseline → 0.0.

Decrease-is-better KRs

score = (baseline − actual) ÷ (baseline − target)

Binary / milestone KRs

| Outcome | Score | |---|---| | Fully achieved | 1.0 | | Partially achieved | 0.5–0.7 per profile; default 0.5 | | Not achieved | 0.0 |

Commit vs aspirational at scoring time

| Type | How to read the score | |---|---| | Commit | 0.7–1.0 = delivered. Below 0.7 = miss worth explaining. | | Aspirational | 0.6–0.7 = healthy stretch [unverified — common OKR convention]. 1.0 = overperformance or sandbag — check calibration. |

Workflow

Step 1: Load profiles

Read org-profile.md and okr/CLAUDE.md before scoring.

Step 2: Score each KR

State baseline, target, actual, formula, score, commit/aspirational label. Show arithmetic.

Step 3: Roll up to objective — do not average

Use decision tree: Achieved | Partially achieved | Not achieved. Mandatory divergence checks:

1. KRs scored well, objective wasn't served.
2. KRs missed, objective substantially advanced.

Step 4: Next-cycle recommendation per objective

keep | kill | revise | promote-to-commit — "keep" should be least common.

Step 5: Cross-team patterns

Shared external causes, sandbagging, systematic proxy failure.

Step 6: Gaming-pattern check before output.

Output format

CONFIDENCE: [defensible recommendation | structured first pass]
SCORING FORMULA: [from profile, or default]

OBJECTIVE: [text]
KR scores:
  [KR text]: [score] ([commit | aspirational]) — baseline [x] → target [y], actual [z]

OBJECTIVE-LEVEL CALL: [achieved | partially achieved | not achieved] — [one line]

DIVERGENCE FLAG: [none | KRs hit / objective missed | KRs missed / objective advanced]

NEXT CYCLE: [keep | kill | revise | promote-to-commit] — [why]

[repeat per objective]

CROSS-TEAM PATTERNS (if applicable): [...]
CALIBRATION NOTES: [...]

Worked example

Input: Objective "Win mid-market." KR "ARR $10M" scored 1.0 via one-off deal; KR "Logo retention 90%" scored 0.4.

Expected output (excerpt):

OBJECTIVE-LEVEL CALL: partially achieved — revenue spike not repeatable
DIVERGENCE FLAG: KRs hit / objective missed — ARR via one-off doesn't prove mid-market win [review]
NEXT CYCLE: revise — replace ARR KR with retention-led measures

Quality checks before delivering

• [ ] Profile loaded or [PROVISIONAL] tagged
• [ ] Every KR has formula and arithmetic or INPUT NEEDED
• [ ] Objective calls do not default to KR averages
• [ ] Divergence types checked per objective
• [ ] "Keep" justified, not assumed
• [ ] Sandbagging signals named if present
• [ ] Gaming-pattern check run

Propose profile update

When a stable convention surfaces during this run (thresholds, naming, tone, output format, or recurring corrections), propose a profile update: show the exact diff against ~/.claude/plugins/config/claude-for-strategy/okr/CLAUDE.md (org-wide facts go to org-profile.md), ask for confirmation, and write only on yes. Only /okr:practice-setup auto-applies a full profile write.

Outputs

Follows plugin CLAUDE.md § Outputs. Next: set-targets for revised KRs, calibration session, or draft-objectives for kill/revise.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: daddia
• Source: daddia/claude-for-strategy
• License: MIT

Install and usage instructions live in the source repository linked above.