create-strategy: interview through the kernel, then write STRATEGY.md

STRATEGY.md is the 12+ month, market-driven layer: given finite capacity, which jobs are funded this period and which are explicitly deferred. Its spine is Rumelt's kernel: Situation (diagnosis) → Strategy (guiding policy) → Pillars (coherent action) → Why This Wins (the argument). It is a versioned artefact: the git diff shows what changed, decision records show why (see Strategy as Code).

The one rule of this skill: every claim earns its tag, and the strategy survives the smell test before it is written. Most bad strategy is not wrong, it is fluff: goals restated as plans, buzzwords carrying claims they cannot cash, a list of everything with nothing deferred. The interview exists to squeeze that out before it reaches disk.

[!NOTE]

This skill authors a new strategy (a new product, or a new period superseding the old). Changing an active strategy mid-period is a ratified event that goes through amend-durable-doc with a decision record.

Step 0: Ground yourself before asking anything

• The anchors: strategy references the vision and must not fund anything that crosses an invariant. The anchors stay stable while strategy changes; do not blur the layers.
• The Product Spec and Job Specs: pillars fund Job Specs by slug; know what exists so the strategy links rather than re-describes.
• The previous STRATEGY.md and its decision records, if any: a new period responds to the last one; deferrals and data gaps carry forward or close.
• The template (templates/strategy.md) and its embedded bad-strategy smell test: your output shape and your final gate.

Step 1: Interview through the kernel (one question at a time)

Ask one question, wait, reflect back, then ask the next. Do not draft until every part is settled.

• The challenge. What is the specific, hard obstacle this strategy faces? "Grow revenue" is a goal, not a challenge. If the user answers with a target, ask what stands between them and it; that is the diagnosis. No crisp challenge, no strategy; say so.
• The Situation, claim by claim. For each claim about the market, the competition, or the customer: how do you know? The answer sets the tag: data you can point to is [evidenced] (get the link or the source name); a reasoned belief is [hypothesis] (record what it is based on); an open question is [to-validate] and goes to Data Gaps with a validation path. Never let a hypothesis wear an evidenced tag because it is confidently phrased: the tags tell readers what is load-bearing, and a mislabeled one quietly transfers risk to everyone downstream. Keep only load-bearing claims; a Situation that pads reads as diagnosis but decides nothing.
• The thesis. Given that situation, what is the overall move, in one paragraph of plain language? This is a guiding policy, an approach with a point of leverage, not a restated target. Test it: does it tell you what to say no to? A thesis that rejects nothing is fluff.
• The allocation. Where does capacity go, in percentages that sum to 100? The percentages force the trade-off to be explicit; a strategy that will not commit numbers is hedging. Include the unglamorous bucket (debt / resilience) deliberately rather than by omission.
• The decision tests. What must be true of an initiative for it to be funded this period? Then the check: would each proposed initiative the user currently has in mind pass? If everything passes, the tests are not doing any work; tighten them until they reject something real, and name what they reject.
• The pillars. For each funded bet: the thesis (what reality it responds to), the observable outcome or measure, and the Job Specs it funds, by slug. If a pillar funds a job with no Job Spec, that is a named gap for create-job-spec, not a reason to invent one. Each pillar maps to a row in the allocation table.
• The deferrals, explicitly. What are you deliberately NOT doing, why not now, and what would change it? This is the section users most resist and the one that makes the strategy legible. A strategy with an empty deferral table has not made choices; do not leave this section until it has real entries someone will be unhappy about.
• The data gaps. Every hypothesis and to-validate claim from the Situation lands here with: what would confirm it, what would falsify it, and how to get it. A gap without a falsification condition is not a gap, it is a hope.
• Why this wins. The coherent-action argument: why this set of bets with this capacity produces the outcome, rather than a different allocation of the same capacity. Ask for the dependencies it rests on and what would break it. It must survive an adversarial read; write it so a reviewer can find the weak points, because one will look.

Step 2: Run the bad-strategy smell test, then write

Before writing, walk the template's four checks with the user, honestly:

• No fluff -- every Situation and Strategy line survives a plain-language rewrite.
• Faces the challenge -- the Situation names a specific, hard obstacle.
• Strategy, not a goal -- the Strategy section gives a chosen approach, not just a target.
• Clean objectives -- the pillars are a focused few, each more achievable than the raw challenge.

If any line fails, go back to the interview; the file is not yet a strategy. Then fill templates/strategy.md exactly: period and status in the header, tagged Situation, thesis, allocation table, decision tests, pillars with funded Job Spec slugs, deferrals, data gaps, Why This Wins, and the smell test checked at the bottom.

If this strategy supersedes a previous one, mark the old file superseded and write the accompanying decision record (context, decision, consequences) in decisions/, linking the commit.

Step 3: Hand it to the gate

Hand the draft to review-doc as a fresh process. The reviewer will hit the tags, the empty-deferral smell, and the pillar → Job Spec links; iterate to APPROVE, then the strategy goes to the humans who ratify it. Strategy is ratified by people, not by agents: your job ends at a clean, honest draft.

Output

1. STRATEGY.md, written to the strategy directory (confirm the path), plus a decision record in decisions/ if this supersedes an active strategy.
2. The claim ledger: every Situation claim with its tag and source, so the ratifying humans can see what is load-bearing at a glance.
3. The gaps: Job Specs the pillars fund that do not exist yet, and the data gaps with their validation paths, as an honest backlog.
4. The review-doc verdict.

The self-teaching check (dogfood this skill)

Give a fresh agent this skill, the template, and a user whose opening ask is "our strategy is to grow ARR 3x by being the best platform". It passes if it refuses that as a goal-not-strategy, interviews to a specific challenge, tags every claim honestly (downgrading confident-but-sourceless ones), forces real deferrals, and runs the smell test before writing. If the output's Situation contains an untagged claim or the deferral table is empty, it failed.

Related

• Strategy template -- the output shape, including the smell test
• Strategy as Code -- the method: versioned strategy plus decision records
• Decision Record template -- written when a strategy is superseded or materially changed
• amend-durable-doc -- the guarded path for changing an active strategy
• review-doc -- the fresh-process gate this draft exits through