Product support

You need to be in a task inside your Project, not a loose chat. Make sure you created the Project first (Cowork → Projects → New Project), opened it, and started the task there. The wizard refuses to run without a Project folder mounted — that's by design, because the Project is where your profile is saved.

This is the failure mode to watch most closely. Two fixes:

• Verify every IRC section, Reg, Rev. Proc., and dollar threshold against current authority before anything goes to a client. tax-specificity-guard flags drafts that state a citation precisely so you don't miss one.
• When you catch a fabricated cite, paste it back and ask: "Where did you get this citation? I didn't provide it." Then add to your request: "Never cite authority I haven't provided — use [verify citation] as a placeholder instead."

Re-run /accountant-setup-wizard and list those exact phrases under "words I never use" — every skill reads them. If one keeps slipping through, add a hard line to your request: "Do not use the word 'write-off' in any client-facing document."

Add to your request: "Include 'as of [TAX YEAR]' on any reference to current tax law, rates, or thresholds." A planning memo without it becomes a problem when the client reuses it next year.

• Your Project folder is probably thin. Add 3–5 of your best anonymized client letters so Claude can see your actual voice, then re-run /accountant-setup-wizard to sharpen the voice answers.
• Or the request didn't ask for enough specificity. Re-run with: "Match the voice from my sample letters in the Project folder, and reference the specifics of [this client's situation]."

Add an explicit limit: "150 words max." If Claude keeps going over, add: "When I give a word limit, stay under it — cut content rather than exceed it."

That's practice-boundary-guard doing its job — it fires when a request edges into legal advice or work outside the engagement. Re-scope the request to what you're actually engaged to do, or send the client a scope-change note (/scope-change-notification).

You need to be in a task inside your Project, not a loose chat. Make sure you created the Project first (Cowork → Projects → New Project), opened it, and started the task there. The wizard refuses to run without a Project folder mounted — that's by design, because the Project is where your profile is saved.

You don't answer it — you route it. Run /tax-or-bookkeeping-triage (acknowledge and send to their CPA) or /refuse-tax-advice (a friendly decline). This is the most important habit in the whole product; the cpa-boundary-guard is there to catch it if you forget.

No, that's cpa-boundary-guard doing its job. It fires when a request drifts into tax advice, worker classification, payroll-tax, sales-tax/nexus, audit language, or entity choice. Re-scope to the bookkeeping side (record the facts, route the determination), and you're good.

client-funds-control-guard reframes "what to pay" into "what the client should review and approve." You prepare the list; the client or an authorized approver decides. Stay out of custody risk.

• Your Project folder is probably thin. Add 3–5 of your best anonymized client emails, then re-run /bookkeeper-setup-wizard to sharpen the voice answers.
• Or the request didn't ask for specificity. Re-run with: "Match the voice from my sample emails in the Project folder, for [this client's situation]."

That's exactly what tax-specificity-guard flags — verify it before sending. Never let a stated number go to a client unverified; use [verify current rule] until you've confirmed it.

Add an explicit limit: "150 words max." If it keeps going over, add: "When I give a word limit, stay under it — cut content rather than exceed it."

That's the one thing to watch, and the skills are built to avoid it — but if you see a number or feature you didn't provide, replace it. The rule the pack follows: if you didn't supply it, it becomes a bracketed placeholder, not a guess. For comps specifically, the pack has no live MLS access — always pull the comps yourself and paste them in. realtor-practice-boundary-guard flags value and comp claims so they don't slip by.

The fair-housing-guard runs automatically on anything headed for a client, the MLS, or the public, and appends a before-publishing checklist. Trust it, but still do a final read with one rule in mind: describe the property and the location, never the buyer. If you're marketing a 55+ community, the guard will only allow age-targeted language with a "confirm valid 55+ status for this listing" stamp — fill that in from the actual listing, never from your specialty alone.

The pack won't answer those, by design — realtor-practice-boundary-guard reframes them to agent-supplied / professional-reviewed. Draft the part that's yours (the plain-language update, the next step) and route the legal/tax/financing question to the attorney, CPA, or lender. That's not a limitation; it's what keeps you inside your license.

Stop. The pack never originates, relays, or confirms wire instructions, and neither should an email it drafts. realtor-privacy-wire-guard flags this every time. Tell the client to use the verified instructions they already received from title/escrow and to confirm them by phone using a number they independently know before sending a cent. Wire fraud is the biggest loss in this business — this is the one place to be slow.

That's the one thing to watch, and the skills are built to avoid it — but if you see a number or feature you didn't provide, replace it. The rule the pack follows: if you didn't supply it, it becomes a bracketed placeholder, not a guess. For comps specifically, the pack has no live MLS access — always pull the comps yourself and paste them in. realtor-practice-boundary-guard flags value and comp claims so they don't slip by.

The fair-housing-guard runs automatically on anything headed for a client, the MLS, or the public, and appends a before-publishing checklist. Trust it, but still do a final read with one rule in mind: describe the property and the location, never the buyer. If you're marketing a 55+ community, the guard will only allow age-targeted language with a "confirm valid 55+ status for this listing" stamp — fill that in from the actual listing, never from your specialty alone.

The pack won't answer those, by design — realtor-practice-boundary-guard reframes them to agent-supplied / professional-reviewed. Draft the part that's yours (the plain-language update, the next step) and route the legal/tax/financing question to the attorney, CPA, or lender. That's not a limitation; it's what keeps you inside your license.

Stop. The pack never originates, relays, or confirms wire instructions, and neither should an email it drafts. realtor-privacy-wire-guard flags this every time. Tell the client to use the verified instructions they already received from title/escrow and to confirm them by phone using a number they independently know before sending a cent. Wire fraud is the biggest loss in this business — this is the one place to be slow.

You need to be in a task inside your Project, not a loose chat. Make sure you created the Project first (Cowork → Projects → New Project), opened it, and started the task there. The wizard refuses to run without a Project folder mounted — that's by design, because the Project is where your profile is saved.

It shouldn't, and the design prevents it — but if a number appears, never send it. Rates, APR, payments, points, and lock terms always come out as [from engine] / [from LE] placeholders you fill from your pricing engine and the Loan Estimate. If a number slips in, replace it with a placeholder and add to your request: "Never state a rate, APR, payment, or lock term — use [from engine] placeholders only." lo-compliance-guard flags drafts that drift toward a quote.

That's a UDAAP trap — stop. Add to your request: "No future-rate predictions, no urgency or scarcity claims, no guaranteed-savings language." The guard flags these, but the judgment is yours.

The pack drafts the copy; RESPA Section 8 compliance is yours. Confirm the cost split reflects fair market value and you've documented who paid for what before publishing. If you're unsure whether a piece crosses the line, send it to your compliance officer first — Section 8 violations carry per-occurrence penalties.

Never describe a borrower by a protected class. Add to your request: "Describe the loan, the property, and the math — never the borrower's protected characteristics." Re-run and scrub before sending.

• Your Project folder is probably thin. Add 2–3 of your best real borrower or realtor emails (scrubbed of names and NPI) so Claude can see your actual voice, then re-run /lo-us-setup-wizard to sharpen the voice answers.
• Or the request didn't ask for enough specificity. Re-run with: "Match the voice from my sample emails in the Project folder, and reference the specifics of [this borrower's situation]."

Add an explicit limit: "100 words max." If Claude keeps going over, add: "When I give a word limit, stay under it — cut content rather than exceed it."

Use /denial-with-options for the borrower communication, but the formal adverse-action notice still goes through your compliant process — it has ECOA timing and content rules the guard flags. The draft is the message, not the legal notice.

/lo-us-appraisal-update communicates appraisal status (ordered / received / delayed) and states no value. It does not negotiate value, advise on a reconsideration-of-value (ROV), or counsel the contract — that's a live conversation with the borrower and agent, not an AI draft.

You need to be in a task inside your Project, not a loose chat. Make sure you created the Project first (Cowork → Projects → New Project), opened it, and started the task there. The wizard refuses to run without a Project folder mounted — that's by design, because the Project is where your profile is saved.

This is the failure mode to watch most closely. Two fixes:

• Verify every code citation, permit requirement, and inspection rule against your local building department or your reference materials before anything goes to a customer or the city. Codes are local and they change.
• When you catch a fabricated reference, paste it back and ask: "Where did you get this code citation? I didn't provide it." Then add to your request: "Never cite a building-code section I haven't provided — use [verify with local code] as a placeholder instead."

Always check the math yourself — Claude is not a calculator and doesn't know your market. Give it your actual numbers in the prompt; don't ask it to estimate prices for you. On change orders and estimates it's built to mark [verify total] rather than calculate — keep it that way.

That's the warning sign to stop and route it to your attorney. The pack is explicit that contract, lien, and waiver instruments come from counsel — lien-and-license-guard flags these and appends a checklist, but the actual legal step is the attorney's, and a statutory deadline is yours to track. Never let a lien or notice window slip.

Add to your request: "Use 'estimated,' 'approximately,' and 'subject to site conditions.' Never write 'guaranteed' about price, timeline, or outcome." The advertising-environmental-scope-guard also flags guarantee language and "licensed/insured/bonded" claims for substantiation.

• Your Project folder is probably thin. Add 2–3 of your best real customer emails (scrubbed of names and numbers) so Claude can see your actual voice, then re-run /contractor-setup-wizard to sharpen the voice answers.
• Add the phrases you'd never say to the "words I never use" list, and the phrases that sound like you to the voice samples.

Add them to your "words I never use" list, and add a hard line to the request: "Use plain language. No fancy construction or corporate jargon."

• Your Project context is probably empty. Add 2–3 real estimates and emails (scrubbed).
• Your trade and specialty in the setup are probably too broad. Get specific.
• Add specificity to the request: "This is for a [specific job type] customer in [specific situation]."

For an everyday field change to scope, cost, or schedule, use /change-order. Only use /contract-addendum-cover-letter when you're routing a formal, attorney-prepared amendment to the contract. When in doubt, it's /change-order.

That's advertising-environmental-scope-guard or lien-and-license-guard doing its job — they fire when a request edges into legal advice, engineering, specialty-license work, or another trade. Re-scope the request to what you're licensed to do, and route the rest to the right professional.

You need to be in a task inside your Project, not a loose chat. Make sure you created the Project first (Cowork → Projects → New Project), opened it, and started the task there. The wizard refuses to run without a Project folder mounted — that's by design, because the Project is where your profile is saved.

• Re-run /photographer-setup-wizard and add them to your "words I never use" list explicitly. Every skill reads it, and photographer-voice-keeper enforces it. Common bans: "capturing memories," "freeze the moment," "tell your story," "magical," "epic," "obsessed," "dream session," "your forever."
• At the end of a prompt you can add: "Don't use any of the clichés from my voice list."

• Your Project folder is probably thin. Add 3-5 of your best real inquiry replies, a gallery delivery email, and a few captions (scrubbed of names) so Claude can see your actual voice, then re-run /photographer-setup-wizard to point at them.
• Add to the prompt: "This must sound specifically like me, not a generic photographer. If any studio could have sent it, rewrite."

• Upload more of your best converting replies — voice comes from your samples, not defaults.
• Always provide the specific detail from their inquiry (venue, date, what they asked) so the reply can be specific to them.

• Add an explicit limit: "100 words max." If Claude keeps going over, add: "When I give a word limit, stay under it — cut content rather than exceed it."

It shouldn't — the design prevents it. If a number appears that you didn't provide, never send it. Replace it with a placeholder and add: "Never state a price, date, or deliverable count I didn't provide — use a placeholder." When you catch one, ask "where did you get this number?" and reinforce the rule.

Add: "All contract-clause output must begin with 'DRAFT — REVIEW WITH ATTORNEY IN [STATE] BEFORE SENDING' and end with a list of points to flag." Have one attorney review all your clauses once — especially the AI-edit, AI-training, and usage-rights language.

No — that's fabrication and breaks platform rules. /review-testimonial-request asks the client for an honest review and makes it easy. If a client shared real words you want to display, the pack tidies their words but stamps it a draft for them to approve and post themselves — never a quotable testimonial it invented.

/reschedule-cancellation-reply drafts the human message. It points to your contract's reschedule/retainer/cancellation terms as you supply them — it never invents a policy. When money is in dispute, review against the signed agreement before sending.

/final-balance-reminder drafts the courtesy reminder. The amount and due date come from your signed contract as placeholders — if you don't supply them, the draft shows placeholders rather than a guess.

Vendor pitches are about a specific shared relationship. Always provide the specific shared client, venue, or context in the prompt — a pitch with no specific shared context will always sound generic. Use /vendor-pitch-wedding or /vendor-pitch-brand.

You need to be in a task inside your Project, not a loose chat. Make sure you created the Project first (Cowork → Projects → New Project), opened it, and started the task there. The wizard refuses to run without a Project folder mounted — that's by design, because the Project is where your profile is saved.

Run skills inside the Project where you ran the wizard — not a free-floating chat. Every skill auto-loads your profile file from the Project folder, so it only works in that Project. If you started a chat outside it, move into the Project and try again.

Three likely causes:

• The cartridge isn't specific enough yet → re-run /cartridge-builder after a couple weeks of real posts.
• You didn't tell Claude which client this is, so it didn't pull the right cartridge → name the client in your request.
• Add "never use corporate jargon" to that client's cartridge never-use list as a blanket rule.

• The cartridge is too thin → re-run /cartridge-builder to refine it, and add 3–5 real sample posts from this client so Claude can see the actual voice.
• Or the prompt didn't ask for enough specificity → re-run with: "This is [Client]. Match the style of their cartridge's sample posts closely."

TikTok wants a different modulation — more casual, faster. Add platform-specific notes to the cartridge describing how the voice shifts per platform, or re-run /caption-tiktok with "make this more casual / faster-paced than the Instagram version."

This is the failure mode that loses client confidence fastest — prevent it obsessively:

• Work one client per request, and never ask for two clients' posts in the same request.
• Name the client so Claude pulls only that cartridge.
• Run /voice-check to compare a draft against the right cartridge, and use the 60-second voice-bleed scan in cheat-sheet.pdf before anything from a different client goes out.

After the first run you should be editing for under two minutes, not ten. If it's more, the cartridge isn't aligned with reality (re-run /cartridge-builder) or the AI is defaulting to generic language (add more real sample posts to the cartridge). Tighten the feedback loop — see why-this-works.pdf.

Your /content-calendar context didn't specify distribution. Re-run with: "Mix pillars throughout — no more than 2 consecutive posts in the same pillar." If the imbalance persists, review the cartridge's pillars with /cartridge-builder.

The pack is built to use real, attributable sources and to flag uncertain ones rather than invent them — where a specific statute or CFR section is uncertain you'll see "[verify exact citation before client use]". Treat that as a feature: verify any specific citation before it goes to a client. If you ever see a confident-looking citation you can't source, don't ship it — ask the skill where it came from.

The pack won't answer those, by design — practice-boundary-guard reframes anything that strays into a legal/tax/audit/valuation/regulatory determination to a recommendation for the client's licensed advisor. Draft the part that's yours (the strategy, the options, the questions for counsel) and route the determination to the right professional. That's not a limitation; it's what keeps you inside your scope of practice.

Use /linkedin-post-from-engagement — it's built NDA-safe (no client names, no deal specifics), and confidentiality-guard double-checks it. The rule it enforces: share the lesson, never the client. If you paste client-identifying details, the guard will flag them with a redaction checklist before anything is public-ready.

If there's any chance of a conflict, run /conflict-of-interest-check before you accept — it walks a structured COI check so you catch the problem before you're committed.