decision-analysis

Decision Analysis Protocol

EXTREMELY_IMPORTANT: This is a MANDATORY protocol, not a suggestion. Follow every step. Do not skip steps. Do not combine steps. Do not summarize. Work through each gate in order.

You're choosing between alternatives. Do NOT default to the first reasonable option. Work through this protocol to find the best option — or discover the question is wrong.

STEP 1 — Frame the real decision

State the decision in one sentence. Then check:

• Are these actually the only options? What's the third option nobody mentioned?
• Is this the right decision to be making? (Sometimes the meta-question matters more)
• Is this reversible (Type 2) or irreversible (Type 1)?

Type 2 → Spend 10 minutes on this protocol. Decide and iterate. Type 1 → Spend proportional effort. Get it right.

STEP 2 — Enumerate options

List ALL viable options, including:

• The options the user named
• "Do nothing" (always a valid option — what happens if we don't decide?)
• The hybrid/phased approach (is there a way to get 80% of the benefit now?)

For each option, state in one sentence what it optimizes for.

STEP 3 — Define criteria and weight them

List 4-6 criteria that matter for this decision. Assign weights (1-5).

Common criteria: cost, speed to deliver, risk, maintainability, team capability, scalability, user impact, reversibility.

Ask the user which criteria matter most. Do NOT assume.

STEP 4 — Score and calculate

For each option, score each criterion (1-10). Multiply by weight. Sum.

Present as a table:

Option	Criterion A (×W)	Criterion B (×W)	...	Total

If two options are within 10% of each other, the decision matrix alone is insufficient. Proceed to Step 5 for tiebreaking.

STEP 5 — Second-order analysis

For the top 1-2 options, apply:

Second-order thinking: What happens 6 months after we choose this? What's the 2nd-order consequence? The 3rd-order?

Pre-mortem: Assume we chose this option and it failed badly in 6 months. What went wrong? Is that failure preventable?

Steelman the loser: Make the best possible case for the option you're NOT recommending. If the steelmanned case is compelling, reconsider.

STEP 6 — Recommend with confidence

State your recommendation explicitly:

RECOMMENDATION: [Option X]
├── Why: [1-2 sentence justification tied to highest-weighted criteria]
├── Key risk: [the most likely way this goes wrong]
├── Mitigation: [how to address that risk]
├── Reversibility: [Type 1/2, what undo looks like]
└── Confidence: [HIGH / MEDIUM / LOW]

If confidence is LOW, state what additional information would increase it, and recommend gathering that information before committing.

STEP 7 — Journal the decision, calibrate confidence

Stated confidence means nothing if nobody ever checks it against reality. Close that loop. If journaling is impractical here (no project workspace, one-off headless task), say so and skip this step — never block a recommendation on it.

7a — Append. Add one JSON line to .praxis/decisions.jsonl in the project root (create it if missing). One tool call, minimal schema:

{"date":"2026-07-03","decision":"<one sentence>","chosen_option":"X","confidence":"MEDIUM","key_risk":"<from STEP 6>","revisit_by":"2026-09-01","outcome":null}

revisit_by is a date or concrete trigger ("after v2 ships") for checking how the decision turned out. outcome stays null until known.

7b — Revisit nudge. If .praxis/decisions.jsonl exists, check it early (at STEP 1, before recommending) for entries with outcome: null whose revisit_by has passed. Tell the user: "N past decisions are due for a confidence check." This is a nudge only — old unresolved entries NEVER block the current decision.

7c — Calibrate. When a journaled decision's outcome becomes known (the user reports back, or you observe the result), set that entry's outcome to "held" / "failed" plus one sentence on what happened — the only permitted edit to past lines. Then compare stated confidence against results: are HIGH calls holding most of the time? Did LOW calls fail the way key_risk predicted? Use the pattern to adjust FUTURE confidence estimates. This is calibration for better future calls, not scoring of past ones.

Do NOT recommend an option without: - At least 3 options evaluated (including "do nothing") - Explicit criteria with weights confirmed by the user - Second-order consequences considered for the recommendation - Confidence level stated

STEP 7 journaling is best-effort, not a gate: if .praxis/ doesn't exist or journaling isn't practical, skip it gracefully — never withhold a recommendation over it.

This skill prevents:

• "Let's just go with X" without evaluating alternatives
• Anchoring on the first option considered
• Ignoring "do nothing" as a valid choice
• Making irreversible decisions with reversible-decision rigor
• Confidence levels that are never checked against outcomes

Superpowers handoff

After recommendation is made and user accepts:

If Superpowers is installed and the decision leads to building something → invoke Skill(superpowers:brainstorming) with the chosen option, criteria weights, and key risks as context. Let Superpowers drive the design of the chosen approach.

If Superpowers is NOT installed → proceed to design or implementation based on the recommendation.