Full SKILL.md
Problem Framing
Turns a vague mandate into a problem statement, MECE issue tree, prioritised branches, hypothesis register, day-one answer and analysis workplan. Use at the start of any engagement or analysis when the brief is a symptom, a wish or a solution in disguise, or when the user asks for an issue tree, problem statement, scope, workplan, "where do we start", "is this MECE", or "the CEO already has a theory". Not for testing hypotheses against evidence (hypothesis-stress-testing), choosing between strategic options (strategic-choice), or product-specific where-to-play questions (product-strategy).
Produces the framing pack a partner expects at the end of week one: SCQ and problem statement, issue tree, prioritised branches with prunes recorded, hypothesis register, day-one answer, workplan. Runs steps 1 to 4 of Conn and McLean's seven-step cycle with Minto's SCQ inside step 1 and Rasiel's MECE and initial-hypothesis disciplines. The reasoning behind each step, and where the sources disagree, is in references/frameworks.md; read it only when a step feels arbitrary or the user challenges the method.
Routing
Three skills answer questions that sound alike. Route by the object the question is asked of:
If the decision-maker cannot be identified at all, say so and ask before building anything.
Inputs
Ask for, or infer and label as assumptions: the mandate verbatim and who gave it; who decides, who recommends, who signs off numbers; deadline and plan horizon; the two or three numbers that prove the complication; constraints (what the client will not do) and positions (what stakeholders already believe); any prior work (SOW, board paper, earlier deck).
Ask one round of clarifying questions if the decision-maker, the target or the deadline is missing. Otherwise proceed on labelled assumptions; a pack built on stated assumptions is easier to correct than a list of questions.
Workflow
Each stage has a "done when" test. Stop at the first stage that fails it and fix before moving on.
-
Find the question. Write the mandate verbatim, classify each clause as Situation, Complication, Question or Answer, then rewrite as SCQ. Situation: undisputable. Complication: contains a number. Question: one sentence in the decision-maker's terms, does not presuppose the answer, carries the target (ask for one or propose one as an assumption). Done when the decision-maker would say "yes, that is what I need to know", not "and also".
-
Fill the worksheet (
references/templates.md§2). Name decider and recommender separately. Success criteria measurable with a timing split. Constraints are decisions, not preferences; preferences go in Known positions. Required accuracy in ± terms (board decisions rarely need better than ±20%). Done when an analyst who has never met the client could read the page and know what to analyse, what to ignore and how precise to be. -
Decompose. Diagnostic tree first ("why is X happening?") if the cause is unknown or disputed; solution tree ("how do we achieve Y?") only when the cause is agreed. Cut level one on a single dimension, preferring an arithmetic identity (profit = revenue × margin − cost; volume = customers × frequency × basket). Fall back to an exhaustive category, then process stage, then time. Segments only below level one. Frameworks (4Ps, Five Forces) check completeness afterwards; they are not branches. Two or three levels; stop when a node is a single analysis someone could run. Quantitative trees show contributions summing to the parent. Done when any client fact lands in exactly one bottom node.
-
MECE tests at every level (
references/templates.md§4). Addition (residual under 5%), classification (one fact, one branch), what-else (three parent contents each land somewhere), count (2 to 5 branches), same-cut (one dimension per level). Strict at levels one and two; approximate below, written down. An "other" branch is acceptable below level one if sized and under 10%. Done when every level has pass or a recorded fix. -
So-what, prioritise, prune. Size each bottom branch to one significant figure in the unit of the success criterion. Ask: if this analysis returned its most extreme plausible answer, what would the decision-maker do differently? "Nothing" means prune regardless of size. Judge ability to influence inside the worksheet's constraints. Place on impact × influence: high/high work first; high impact/low influence watch; low impact prune. Record a reason for every prune. Prioritised branches should sum to 120 to 150% of the target; if they fall short, tell the client now whether the tree is missing an option or the target is unrealistic. Done when every branch has a decision and a reason and the prioritised total exceeds the target.
-
Hypotheses and day-one answer. One hypothesis per prioritised branch, four parts: assertion with magnitude and condition; two to four load-bearing assumptions (flagged, not tested); the observation that would refute it; the branch it came from. A stakeholder prior the tree does not support becomes a negative hypothesis. Then the day-one answer in SCR form, under 120 words, labelled as a hypothesis with a confidence level, two or three gaps and a refresh date. Done when no hypothesis is a topic ("pricing") rather than an assertion ("pricing is worth at least $5M if…").
-
Workplan. One row per analysis: issue, hypothesis, analysis needed (output, not activity), source (an extract or a named person, never "client data"), owner, due. Detail two weeks, outline the rest. Add a week-one row confirming the rewritten question with the decision-maker, a day-one-answer refresh within two weeks, a midpoint steering committee, a draft deliverable a week before the deadline, and a synthesis row that checks initiatives do not double-count. Done when an analyst could start any row tomorrow without asking a question.
-
Package. Assemble in the order below. Method vocabulary stays out of the pack: no stage numbers, checkpoints, "MECE", skill names, framework authors or hand-off notes. One closing sentence on what happens next, in the user's terms.
Output
The pack, in this order and at these lengths, 2,000 to 3,500 words including tables:
- SCQ (3 to 5 sentences)
- Problem statement (under 200 words; worksheet table)
- Issue tree (one page, basis of each split stated)
- Prioritised branches (table: branch, rough impact, influence H/M/L, so-what, decision, reason) plus a three-to-five-line "Not doing" list with reasons and, for each prune, the observation that would revive it
- Hypothesis register (table: ID, hypothesis, load-bearing assumptions, refuting observation, branch)
- Day-one answer (SCR block with date, confidence, gaps, refresh date)
- Workplan (8 to 15 rows, seven columns) and dated checkpoints
- Where a spend or hire is on the table: the breakeven in one line and the decision rule it implies
Assumptions sit beside the content they support (worksheet field, footnote on a tree number), never as an opening list. If the user's figures do not reconcile, one sentence at the point it matters, then continue with figures as given.
Working notes (MECE check table, matrix workings, partner checklist from references/templates.md §9) appear only if asked for, or if something in them changes a decision; then under one heading, "Working notes", after the pack. Full templates in references/templates.md; full example with arithmetic in references/worked-example.md. For a file, hand the pack to docx or pptx.
Quality bar
Run silently before delivering; fix what fails, do not report the check.
- Question read alone: who decides, what outcome, by when are all present.
- Situation read to a sceptical CFO: nothing disputable.
- Level-one branches: a single named dimension.
- Three random client facts: each lands in exactly one node.
- Quantitative levels sum within 5%; stated totals equal the sum of the rows and use the base the worksheet defines.
- No prioritised branch has a so-what of "nothing".
- Prioritised branches exceed the target, or the client has been told why not.
- Every hypothesis has a refuting observation and a magnitude.
- Every prune has a reason.
- The decision-maker's own suspected cause appears as a labelled hypothesis, not a finding.
- The day-one answer cannot be quoted as "the team recommends".
- Search the pack for "Stage", "MECE", "checkpoint", "hand-off", skill names and the worked example's client; any hit is a leak.
Anti-patterns
- Building the tree around the mandate's smuggled answer. Classify clauses first; the answer goes in Known positions or becomes a negative hypothesis.
- Level one as a topic list ("pricing, costs, customers, people"). Cut on an identity or a single split.
- "The tree is MECE" without a check table.
- Modelling to the nearest $100K before ranking. One significant figure here;
value-driver-modelinglater, for prioritised branches only. - A solution tree when stakeholders still disagree about the cause.
- Workplan rows that are intentions ("review pricing data, owner: team").
- Recording a CEO preference as a constraint.