Decision matrix

Decision Matrix

Overview

A weighted decision matrix (a.k.a. weighted scoring model, Pugh matrix, or trade study) turns a fuzzy "which should I pick?" question into a transparent, defensible number. You list the options (the things you choose between), the criteria (what matters), assign each criterion a weight (how much it matters), score every option on every criterion, and compute a weighted total. The highest total is the recommended choice.

Keywords: decision making, weighted scoring, trade-off analysis, multi-criteria decision analysis (MCDA), Pugh matrix, vendor selection, job offer comparison, prioritization.

Use this skill whenever someone must choose among 2+ discrete alternatives and wants the reasoning to be explicit rather than a gut feel. Do not force a matrix on a binary yes/no decision or a problem with a single dominant constraint — say so instead.

Workflow

Follow these steps in order. Confirm with the user at steps 1–4 before computing.

1. Frame the decision. State the single question being answered (e.g. "Which CRM should we adopt?"). One decision per matrix. If the user has bundled several decisions, split them.

2. List the options. Gather 2–7 concrete alternatives. Fewer than 2 is not a decision; more than ~7 becomes noisy — shortlist first. Always consider adding a baseline "do nothing / status quo" option when relevant.

3. Elicit the criteria. Ask what factors matter. Aim for 4–8 criteria. Each must be:

   • Distinct (not measuring the same thing twice — avoid double-counting).
   • Discriminating (options actually differ on it; drop criteria where all score equally).
   • Phrased so higher = better (rename "Cost" → "Affordability", or mark it as a cost criterion to invert — see references/scoring-guide.md).

4. Set weights. Assign each criterion a weight reflecting its importance. Use any consistent scale (1–5, or percentages summing to 100). Normalize internally so totals are comparable. Capture must-have constraints separately as pass/fail gates, not weights — a dealbreaker should eliminate an option, not just dock points (see references/scoring-guide.md, "Constraints vs. criteria").

5. Score each option × criterion. Use a fixed scale (default 1–5; 1 = poor, 5 = excellent). Score one criterion across all options at a time (column-wise) to keep scoring calibrated. Note the rationale for extreme scores.

6. Compute. Weighted score = Σ(weight × score) per option, normalized to a 0–100 index. Use scripts/decision_matrix.py to do this reproducibly from a JSON or CSV file rather than doing arithmetic by hand.

7. Sanity-check & sensitivity. Ask: "Does the winner feel right?" If the top two are within ~5%, the decision is a near-tie — flag it and run sensitivity analysis (vary the most subjective weights/scores; see the script's --sensitivity flag). State which assumptions the result hinges on.

8. Report. Present the ranked table, the winner, the margin, the key drivers, and the risks/caveats. Use templates/report.md. Be explicit that the matrix informs the decision — it does not replace judgment.

Scoring Scale (default)

Score	Meaning
5	Excellent / best in class
4	Good / above average
3	Adequate / acceptable
2	Weak / below average
1	Poor / unacceptable on this axis

Keep the scale consistent across all criteria. For details on normalizing mixed units (dollars, days, ratings) and handling cost-style criteria, see references/scoring-guide.md.

Worked Example (abbreviated)

Choosing a project-management tool across Cost, Ease of use, Integrations, Support.

Criterion	Weight	Tool A	Tool B	Tool C
Affordability	3	5	3	2
Ease of use	5	3	5	4
Integrations	4	3	4	5
Support	2	4	3	5
Weighted		47	52	52

Tool B and Tool C tie at 52. That is the signal to stop and look deeper: Tool C wins on the highest-weight-adjacent criteria but loses on cost; Tool B is balanced. Run sensitivity on the Ease-of-use weight. See examples/tool-selection.md for the full walkthrough, including the JSON input and the script output.

Best Practices

• Weight before you score. Set weights and criteria before seeing the options' scores, to avoid rationalizing a pre-chosen winner.
• Score column-wise, one criterion at a time across all options, for calibration.
• Separate gates from scores. Must-haves are pass/fail filters applied first.
• Show your work. Always surface the table and weights, never just the final number.
• Treat near-ties as ties. A 1-point gap on a subjective 1–5 scale is noise.
• Run sensitivity on the squishy inputs — the weights and scores you were least sure of.
• Keep criteria independent to avoid silently double-counting one concern.

Common Pitfalls

• Fudging weights to get the answer you wanted. If you adjust inputs after seeing results, say so explicitly and re-justify.
• Too many criteria. Beyond ~8, low-weight noise drowns the signal. Consolidate.
• Non-discriminating criteria. If every option scores the same, the criterion adds nothing — drop it.
• Mixing cost and benefit directions. Forgetting to invert a "lower is better" criterion silently rewards the worst option. See references/scoring-guide.md.
• False precision. Reporting 73.42 implies accuracy the inputs don't have. Round and present margins, not decimals.
• Ignoring dealbreakers. A weighted average can let a strong option survive a fatal flaw. Apply hard constraints first.

Supporting Files

• references/scoring-guide.md — scales, normalization of mixed units, cost vs. benefit criteria, weighting methods (direct, ranking, pairwise/AHP), and constraints vs. criteria.
• scripts/decision_matrix.py — stdlib Python tool. Computes weighted scores and rankings from JSON/CSV, prints a formatted table, supports cost criteria and --sensitivity.
• examples/tool-selection.md — full worked example with input file and script output.
• templates/report.md — fill-in template for presenting the decision and recommendation.