Reflect

Reflect

Structured reflection at daily, weekly, and quarterly cadence that ends in a changed behaviour rather than a paragraph of feelings. The mechanism is testing recorded beliefs against outcomes: written predictions scored for calibration, and commitments tracked for whether they actually held.

Boundary with weekly-review: that skill runs the weekly operating cadence — what happened, what is next, priorities and blockers. This one is the learning layer on top: was my judgement any good across weeks and quarters, and what should change as a result. Run them back to back, operating review first, using its output as this skill's raw material. Do not duplicate the wins/blockers synthesis here.

When to use this skill

• Your reviews produce pleasant notes but nothing ever changes as a result
• You want to know whether to trust your own confidence when making a call
• Delivery dates keep slipping and you suspect the estimates, not the execution
• The same commitment has been carried for months without progress or a decision
• It is quarter end and you need to score judgement, not just report outcomes
• You are starting a prediction log and want a scoring method rather than a journal

Inputs the skill expects

• A prediction log — statement, confidence, and outcome once resolved
• A commitment log — text, status (kept/missed/partial/open), optional due and carried_cycles
• The cadence you are running: daily, weekly, or quarterly
• A reference date for overdue calculations (passed explicitly — the tools never read the clock)
• Optionally a domain per prediction, which is where the most actionable signal appears

Clarify First

Before generating, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Which cadence is being run — daily, weekly, and quarterly ask genuinely different questions; the wrong set produces either triviality or an hour-long session that gets skipped
• Whether resolved predictions exist — below 20 resolved, calibration is too noisy to act on and the honest output is "keep logging," not a score
• Whether a weekly operating review already runs — determines whether this layers on top or has to carry the operating cadence too
• Carry-count history on open commitments — the three-cycle rule is the sharpest mechanic here and needs the count to fire

Stop rule: ask only the 2-3 that most change the output. If the user says "just draft it," proceed and list your assumptions at the top of the artifact.

Workflows

Workflow 1 — Score prediction calibration

Quarterly. The step that converts reflection from storytelling into measurement.

1. Resolve every prediction past its date — true or false, no revising the original confidence.
2. Run the scorer with 5 buckets (10 buckets need roughly 50+ predictions to be readable).
3. Read the calibration table first: gaps above 10 points in buckets holding 5+ predictions are real bias, not noise.
4. Read the domain breakdown — bias is rarely uniform, and the worst domain is usually your own delivery dates.
5. Pick exactly one correction and re-measure over a full quarter.

python3 personal-productivity/reflect/scripts/calibration_scorer.py \
  --input personal-productivity/reflect/assets/sample_predictions.json \
  --buckets 5

The decomposition is computed over distinct confidence values, so --buckets changes only the displayed table and never the statistics. Verify the identity BS = REL - RES + UNC and the other scoring invariants at any time:

python3 personal-productivity/reflect/scripts/calibration_scorer.py --selftest

Isolate a single domain once you know where the bias lives:

python3 personal-productivity/reflect/scripts/calibration_scorer.py \
  --input personal-productivity/reflect/assets/sample_predictions.json \
  --domain delivery --buckets 5 --format json

Workflow 2 — Generate a cadence-appropriate prompt set

Weekly, or at whichever cadence you are running.

1. Update commitment statuses from the past cycle — kept, missed, partial, open.
2. Increment carried_cycles on anything that rolled over again.
3. Run the generator for the cadence, with today's date for overdue detection.
4. Work the core prompts, then the accountability prompts — each of those needs a decision, not a note.
5. Satisfy the closing requirement. If nothing changed, the session was journaling.

python3 personal-productivity/reflect/scripts/reflection_prompt_generator.py \
  --input personal-productivity/reflect/assets/sample_commitments.json \
  --cadence weekly --as-of 2026-07-21

Quarterly, where structural change is allowed — note the quarterly log carries carried_cycles, which is what fires the three-cycle rule:

python3 personal-productivity/reflect/scripts/reflection_prompt_generator.py \
  --input personal-productivity/reflect/assets/sample_commitments_quarterly.json \
  --cadence quarterly --as-of 2026-07-21 --format json

Workflow 3 — Run the quarterly reflection end to end

90 minutes, once a quarter. The only cadence where role, commitments, and method are on the table.

1. Score the prediction log (20 min) — Workflow 1.
2. Read the quarter's weekly reflections in one sitting (15 min). Individually unremarkable; in a batch they expose patterns invisible at weekly resolution.
3. Work the quarterly prompts from assets/quarterly-reflection-template.md (20 min).
4. Decide (20 min): kill one commitment, change one method, and give every chronic commitment an explicit date/delegate/kill decision.
5. Write next quarter's predictions with confidence numbers (15 min).

python3 personal-productivity/reflect/scripts/reflection_prompt_generator.py \
  --input personal-productivity/reflect/assets/sample_commitments_quarterly.json \
  --cadence quarterly --as-of 2026-07-21

Decision frameworks

Cadence selection

[PROVEN] Start with weekly only. The most common failure is starting daily because it looks smallest, missing three days, and abandoning everything. Weekly carries most of the value and survives a missed week without collapsing.

Reading a Brier score

The score decomposes into reliability (miscalibration — lower better) and resolution (discrimination — higher better). The common pattern is decent reliability with near-zero resolution: you have learned to hedge everything to the base rate, which is safe and useless. The reverse — sharp judgement, wrong numbers — is more valuable, because numeric calibration is easy to correct and directional judgement is not.

Commitment keep rate

The instinct at a low keep rate is to try harder, which reliably fails — the cause is volume, not effort. Halve the commitments and the rate usually recovers on its own.

The three-cycle rule [PROVEN]

A commitment carried three cycles without progress is not waiting for time; it is waiting for a decision you keep declining to make. Carrying it a fourth time is the decision — to never do it — so make it explicitly: commit to a date, delegate it, or kill it.