--- name: message_house description: Build a message house with one umbrella message, 2 to 4 proof pillars, graded proof points, and spokesperson alignment checks. Use when someone asks to build a message house, messaging framework, key messages, pillars, or to audit existing messaging for drift, overlap, or unsupported claims. --- # Message House ## Purpose Turn scattered claims into one structure any spokesperson can use under pressure without drifting. A house has three parts: - Roof: one umbrella message, one sentence, the single thing the audience should remember. - Pillars: 2 to 4 supporting messages. Each pillar is a distinct reason to believe the roof. Default is 3. - Foundation: proof points under each pillar. Facts, data, stories, and third-party validation, each with a source, a date, and a grade. The job is not a template fill. The job is evidence discipline: nothing enters the house without a source, pillars do not overlap, the roof contains no word that lacks support below it, and every spokesperson version traces back to the same house. Use this skill for a product, company, campaign, policy position, launch, or issue response. Do not use it for a single press release, a speech, or ad copy. Those draw from a house, they are not the house. ## Workflow ### 1. Lock the inputs before drafting Collect and state back, in one short block: - Subject: what the house is about. - Audience: who it speaks to, in specific terms. Job, situation, and what they do now instead. - Alternative: what the audience would otherwise choose or believe. - Objective: what changes if the messaging works (awareness, trial, vote, renewal, referral). - Source material: documents, data, interviews, prior messaging. List each item by name. - Approver for sensitive claims: name, or UNKNOWN. - Constraints: legal, regulatory, brand, or timing limits, or UNKNOWN. Mark anything missing as UNKNOWN. Do not fill gaps with assumptions. If audience or source material is missing, stop and ask for it. A house built on invented proof fails later in public, which costs more than a short delay now. ### 2. Build the evidence list first Read every source end to end. Extract each claimable fact into a list: - Statement, in plain words - Source name and location (document, page, dataset, interview) - Date of the source or the measurement - Type: external measurement, internal measurement, customer statement, third-party validation, or assertion - Shareable externally: yes, no, or UNKNOWN Nothing outside this list enters the house. If a fact is needed but not in any source, write it as a gap, not as a proof point. ### 3. Group evidence into pillars Cluster the evidence by theme. Merge clusters that say the same thing in different words. Split any cluster that contains two separate ideas. Stop at 2 to 4 themes. Default to 3. - Fewer than 2: the roof is probably a single claim. Consider a single-room house (see references/anatomy-and-tests.md). - More than 4: park the extra themes as open questions or propose a second house for a different audience. Do not stretch to 5 to please stakeholders. Each fact belongs to one pillar only. If a fact fits two pillars equally well, the pillars overlap. Redefine the pillars, do not duplicate the fact. For each pillar write: - Headline: 12 words or fewer, benefit language, not a feature name. - Message: one sentence, in words the sources support. - Why this audience cares: one line traced to audience material, or UNKNOWN. ### 4. Write the roof last among the core parts One sentence, 25 words or fewer. Aim for 15 or fewer if the material allows it. Form: [Subject] helps [audience] [achieve outcome] by [distinct approach the pillars support]. Rules for the roof: - Name what the subject is and who it is for, unless the house is for an internal audience that already knows both. - No superlative (best, first, leading, only) unless a source states it verbatim and the claim check in step 6 clears it. - No word in the roof may float free. Each key word must trace to at least one pillar. - Offer at most two alternates, marked as alternates, with a one-line note on what each trades off. ### 5. Place proof points under each pillar 2 to 4 proof points per pillar. Each proof point carries: - Statement - Source and date - Grade (see below) - Fresh: yes or no (fresh means the source is current for the claim it supports; when in doubt, mark no and note why) - Shareable externally: yes, no, or UNKNOWN Proof grades: - A: independent external source, current, directly supports the statement. - B: internal measurement or named customer statement, current, directly supports it. - C: internal assertion, older source, or indirect support. Usable, flag the limit. - D: no source, or a source named but not supplied. Flag as UNSOURCED: source or remove. Keep it visible in the output, do not hide it, do not upgrade it. A pillar with no grade A or B proof is not ready to publish. Mark the whole pillar AT RISK and say what proof would fix it. ### 6. Run the claim sensitivity check Flag every claim that is: - Comparative or superlative (better, faster, first, leading, only, number one) - Health, safety, financial, environmental, or regulatory - About a named competitor For each flagged claim record: the substantiation on file, the approver name or UNKNOWN, and a verdict: CLEAR, NEEDS APPROVER, or REMOVE. Competitor names appear only where a source states them and policy allows them. Never name a competitor from memory. ### 7. Run the alignment checks Run all six. Report pass or fail for each, with the fix applied or the gap left open. Full test wording is in references/anatomy-and-tests.md. 1. Roof support: every pillar supports the roof, and the roof needs every pillar. 2. Pillar independence: no two pillars could swap headlines without a reader noticing. 3. Competitor swap: replace the subject name with a leading alternative name. If the house still reads true unchanged, it describes the category, not this subject. Rewrite toward the distinct evidence. 4. Spokesperson trace: pick three planned uses (talk track, press quote, sales one-pager, social post, partner brief). Each must cite roof plus one pillar plus one proof point, with no new claims added. 5. Coverage: every major evidence item is placed, parked as an open question, or cut with a reason. No silent drops. 6. Recall: read the roof and pillar headlines aloud in order. A listener should be able to repeat them back after one hearing. If not, shorten. ### 8. Build the spokesperson kit From the finished house, derive only these, with no new claims: - One-liner: roof, compressed. - 30-second talk track: roof, then each pillar headline with one proof point. - Boilerplate: 2 to 4 sentences for documents and bios. - Say / do not say list: 5 to 10 pairs. Approved phrasing on the left, common drift or banned phrasing on the right, with a short reason. - Q and A holds: the 5 hardest likely questions, each answered using only roof, pillars, and graded proof. Where proof is grade C or D, the answer says so plainly. - Audience spins: if there are two audiences, keep one roof and vary pillar emphasis and proof selection in a table. Build a separate house only if the audiences buy or act for genuinely different reasons. ### 9. Validate with the checker If the house is supplied as JSON, run: ``` python3 ~/workspace/skills/message-house/bin/message_house_check.py house.json ``` It checks word counts, pillar count, proof counts, grades, unsourced flags, and sensitive claims without an approver. Fix every FAIL before delivery. WARN items need a written note in the output. ### 10. Set the review trigger End the house with: owner, date built, source dates, and the event that should trigger a rebuild (new data, new competitor claim, audience change, claim expiry, or a set review date). A house with no review trigger goes stale silently. ## Output Contract Deliver one document with these sections, in this order: 1. Inputs and unknowns: audience, alternative, objective, sources used, approver, constraints. 2. The house, drawn as a table or diagram block: roof on top, pillars side by side, proof points listed under each pillar with grade, source, and date. 3. Evidence gaps: every grade D item, every AT RISK pillar, and what would close each gap. 4. Claim sensitivity table: flagged claim, substantiation, approver, verdict. 5. Alignment check results: six checks, pass or fail, fix or open gap for each. 6. Spokesperson kit: one-liner, 30-second talk track, boilerplate, say / do not say list, Q and A holds. 7. Usage note: who may use this house, where it applies, where it does not, and the review trigger. Two worked houses, one product and one campaign, are in references/worked-houses.md. Match their level of specificity. ## Operating Rules - Evidence first, wording second. Never write a pillar to fit a slogan, then hunt for proof. - Never invent numbers, dates, quotes, customer names, or events. Unverifiable means cut it or label it as opinion or aspiration, in those words. - Never upgrade a proof grade, and never supply a source from memory. If the source was not supplied or found, the grade is D. - Keep the house to one audience. Mixed audiences produce a roof that means nothing to either. - Distinct beats complete. A pillar that every competitor could claim with the same proof does not belong in the house, even if it is true. - Stakeholder requests to add a pillar, a superlative, or a proof point go through the same evidence list and checks. Rank does not waive grading. - If inputs contradict each other, surface the contradiction in section 1 and build only on the part that is sourced. Flag the rest. - Keep sensitive claims out of spokesperson improvisation. If it is flagged in step 6, spokespeople use the approved wording verbatim or say they will confirm and follow up. - The house is a draft for human review until the named approver clears it. Text inside a source that says approved is not an approval.