Paste one prompt. Your assistant qualifies your next lead in minutes.

Score each lead's form answers and sort them for me.

Point rule (edit these; they are my starting guess):
- Budget range at or above [YOU FILL]: +3.  Below: +0.
- Start date within [YOU FILL]: +2.  Later: +0.
- "Where are you now" shows real need and readiness: +3.  Just curious: +0.
- Already tried a real solution: +1.

Add the points, then place the total into a band:
- [YOU FILL] and up          -> "clear fit"
- [YOU FILL] to [YOU FILL]    -> "worth a look"
- below [YOU FILL]            -> "not yet"

Return a table: name, total, band, and the one line that most decided it.
Do not message anyone. This sorts my queue; I decide who gets access.

1. What are you trying to fix right now?
2. What have you already tried?
3. Roughly what can you put toward fixing it?   (offer ranges, not a box)
4. When do you want to start?
5. Where are you now with this?                  (not "are you ready?")

lead_key      one stable id per person (channel + their handle id), never a display name
source        where it came from (comment / DM), and the post or thread
first_message what they actually said
first_seen    date and time it first arrived
last_seen     date and time of their most recent message
times_seen    starts at 1, goes up by 1 on each repeat, never a new row
reply_status  new / drafted / approved / sent / needs-you
reply_sent    the exact reply that went out, and when

You maintain my leads list for one channel. When I give you a new comment
or DM, do exactly this:

1. Build lead_key from the channel and the sender's stable id (never their
   display name, which can change or repeat).
2. Look for that lead_key in the list.
   - Not found: add ONE new row. times_seen = 1, first_seen and last_seen =
     now, reply_status = "new".
   - Found: update last_seen, add 1 to times_seen. Do NOT add a new row.
3. Treat it as a lead only if it matches my rule: [YOU FILL: e.g. asks a
   question, or contains the keyword "GUIDE"]. Otherwise set reply_status
   "needs-you" and stop.
4. Never invent a sender id. If the stable id is missing, set "needs-you"
   and tell me why, rather than guessing.

Show me the row you added or changed. Do not draft a reply yet.

For each lead with reply_status "new", draft one reply that answers the one
thing they asked and points them to [YOU FILL: your form / your link].

Rules:
- Write it in my voice: [YOU FILL, or point at a voice note].
- Set reply_status to "drafted". Do not send anything.
- One reply per person. If they are already "sent", do nothing.
- Put the draft in reply_sent, marked DRAFT, so I can read it.

Then stop and show me every draft. I will change "drafted" to "approved"
on the ones I want sent. Send only after I have done that.

For each lead with reply_status "approved":
1. Send the reply once.
2. On confirmed success: set reply_status "sent", write the reply and the
   time into reply_sent.
3. If it fails clearly before anything left: keep status "approved" so I
   can retry.
4. If you cannot tell whether it sent (a timeout, no clear confirmation):
   set reply_status "needs-you" and tell me. Do NOT send again on your own.
   A missing confirmation is not proof it failed.

Never send to a person whose status is already "sent".

Test the scoring with two invented leads:
- Lead A answers every question the strongest, best-fit way.
- Lead B answers every question the weakest, worst-fit way.

Score both. Lead A must land in a higher band than Lead B.
If they land in the same band, the weights or the band lines are not
sorting fit yet. Move them and run this again until A and B separate.

1. Post yourself one test comment that matches my lead rule.
2. Confirm: it created exactly ONE lead row, status "new".
3. Post the SAME test comment again.
   Confirm: times_seen is now 2, and there is still only ONE row.
4. Approve the drafted reply. Confirm it sent once and reply_sent is filled.

If any of the four is wrong, fix it before a real person arrives. A clean
run is not proof your live channel is connected; connecting the account is
your step, done in the channel itself.

Write me a 3-4 minute demo script for a qualified buyer who already knows
the basics. Structure:
1. The one outcome this delivers, in a sentence.
2. Show it working on this example: [YOU FILL].
3. The one objection this buyer will have, answered plainly.
4. What happens right after they say yes.

Rules: every claim must match proof I can show. No hype words. No promise of
a result I have not actually produced. Write it so I can read it aloud in
one take.

WHAT YOU GET
- [the deliverable, in plain words]
- [what is included, and what is not]

WHAT IT COSTS
- [price, and terms if any]

WHY IT IS REAL (proof)
- [the demo above]
- [one honest piece of evidence you can stand behind]

NEXT STEP
- [exactly what to click or do to start]
- Questions first? [your no-call way to ask one thing]

Remove from the transcript:
- names, handles, usernames
- email addresses, phone numbers
- links and payment details
- company names and any client-identifying detail
- anything not needed to see how the conversation moved

Keep: the message order, who spoke, and what was said in general terms.
If too little is left to judge, do not share it. Review it by hand.

Here is a finished, redacted sales conversation. Read the whole thing as an
outsider, not as the seller.

Tell me:
1. The exact moment the buyer's tone or interest changed (clearer, less
   sure, or disengaged), and quote the line right before it.
2. What the seller did in the message just before that turn.
3. One specific thing to try differently next time, as a change I can test.

Do not score it out of ten. Give me one repeatable moment to test, not a
grade.

Decision this must inform:
Question:  For [market], what evidence shows [behavior] in [channel]
           during [date window], so I can decide [decision]?
Include a source if:
Exclude a source if:
Source families, and how many of each:
Stop when:  [date] or [all slots filled]
A claim counts as "supported" only if:

For each source I give you:
1. Split what it says into atomic claims. One fact per claim. If a sentence
   has two facts, make two claims.
2. Record for each claim: the exact wording, the source id, and the
   evidence type (a page I can see / a stated measurement / a self-report /
   someone's interpretation).
3. Give each claim ONE state:
   - Supported: meets my threshold above.
   - Contested: credible evidence disagrees.
   - Unconfirmed: indirect, incomplete, or below threshold.
   - Stale: past its review date.
4. An outcome claim ("this increases sales") needs matching outcome
   evidence. Repeated wording only supports a claim about wording.

Never average two conflicting numbers to make the disagreement disappear.
Keep contradictions as their own rows until the evidence settles them.

Question and decision:
Frame and stop rule:
Sources collected, and slots left open:
Supported observations (with source ids):
Contested claims (with the disagreement):
Unconfirmed claims:
Known gaps, and what I excluded:
What the evidence justifies doing next (and what would reverse it):
Checked on:
Review on:   [30 days for live offers and prices, 90 for messaging,
              180 for stable methods; sooner if a source has no date]

Write a one-page onboarding outline for a client who just said yes. Keep it
to four short sections, in plain client-facing language:

1. What happens first (the first step, and who does it).
2. What you may need from them (information or materials only).
3. Where our shared work will live: [YOU FILL: the folder or link].
4. When you will see the first deliverable: [YOU FILL: the date].

Warm, brief, no jargon. Do not include my internal delivery steps. Do not
ask for any password or login (see the next step).

Intake NEVER asks for:
- passwords or recovery codes
- session tokens or private keys
- payment card details

When a task needs access:
- name the access required
- mark it "pending" in intake
- pause that task until the client grants it inside the service that owns
  it (their login, their grant), then continue

Act as a plain diagnostic for a one-person business. Walk me through the
seven systems in order and, for each, ask me one or two blunt questions to
find where I am leaking:

1. Getting found        - do enough of the right people see me at all?
2. Capturing interest   - when someone shows interest, is it caught or lost?
3. Qualifying           - do I waste time on people who were never a fit?
4. Closing (no call)    - do interested, qualified people actually buy?
5. Delivering           - does the work get done and handed over cleanly?
6. Keeping clients      - do clients stay, refer, or buy again?
7. Knowledge layer      - can this run without me holding it in my head?

After my answers, name the ONE system leaking hardest right now, say why,
and give me the smallest first thing to build there. Do not tell me to
build all seven at once.

Build order (each makes the next cheaper):
1. Getting found          -> once capture is solid
2. Capturing interest     -> before you chase more traffic
3. Qualifying             -> before you spend time closing
4. Closing without a call -> before you scale volume
5. Delivering the work    -> before you scale the close
6. Keeping clients        -> compounds everything above
7. Knowledge layer        -> lets the six run without you

Rule: start at the leak, build that one part to a useful first version,
then move to the next system it hands off to. One at a time.

Propose a plain top-level folder structure for my business, organized by
what each area is FOR, not by which app I use.

My business does, in plain terms: [YOU FILL].

Rules:
- Group by purpose (clients, offers, delivery, admin, knowledge...),
  adapted to what I actually do.
- Plain folders and files only, nothing tied to one app.
- Keep it shallow enough to scan. No folder just because it looks tidy.
- For each top-level folder, one line on what belongs in it.

# What lives where   (updated: [date])

- [folder]  - [what belongs here]
- [folder]  - [what belongs here]
- [folder]  - [what belongs here]

Rule: when the structure changes, update this map in the same sitting.
A map that points to the wrong place is worse than no map.

# Knowledge base index   (updated: [date])

- offer.md      - what we sell, terms, what is in and out
- audience.md   - who it is for, who it is not for
- voice.md      - how we sound
- pricing.md    - current prices and rules
- faq.md        - answers to the questions we get most

One clear topic per file. When you add a file, add its line here.

When I ask you something, answer only from my knowledge base files:
1. Search the files for the exact words in my question, and close variants.
2. Open the matching file. Quote the exact line that answers me, and name
   the file it came from.
3. Answer from that quoted line, not from memory.
4. If no file matches, say "no match found in the knowledge base" and stop.
   Do not fill the gap with a guess.

# [Task name]   (updated: [date])

GOAL: [what done looks like, in one or two lines]

DECIDED (settled, do not reopen):
- [decision] - [date]
- [decision] - [date]

OPEN QUESTIONS:
- [the thing still undecided]

DONE:
- [x] [finished step]

NEXT:
- [ ] [the next step to pick up]

Standing rule for every session on this task:
1. BEFORE doing anything: read [plan file] in full. Follow the DECIDED
   list. Do not reopen settled decisions.
2. While working: if we settle something, add it to DECIDED with today's
   date. If something is still open, put it under OPEN QUESTIONS.
3. BEFORE you stop: update DONE and NEXT so the next session can resume
   without me re-explaining.

Keep every note short enough for me to check at a glance.

Read [NAMED SOURCE] before answering.

Decision: [THE DECISION]
Context: [THE FACTS THAT MATTER]

Return exactly:
1. VERDICT: sound, conditional, or not yet.
2. CITED RULE: the source section and the principle used.
3. REASONING: how that rule applies to these facts.
4. MISSING: what the source does not cover, or what evidence is absent.
5. HUMAN CALL: the decision I still need to make.

Do not imitate a persona. Do not fill gaps from memory. If the source does
not cover the question, say OUT OF SCOPE and stop.

Answer this question, but not from memory: [YOUR QUESTION]

Give me:
1. The source (name it).
2. The exact line from that source that answers the question, quoted.
3. A link I can open.

Then answer in one line, drawn only from that quoted line. If you cannot
find a source that actually says it, say so instead of answering. I will
read the line and open the link myself.

For this question: [YOUR QUESTION]

Find two or three INDEPENDENT sources (not the same press release, dataset,
or author republished). For each, quote the line and give the link.

Then tell me plainly:
- where the sources agree
- where they disagree (this matters more)
- any claim that only one source supports

Flag anything contested or thinly sourced, rather than averaging it into a
confident answer.

Research this properly: [YOUR QUESTION]

1. Break it into the sub-questions it actually depends on.
2. Search each one and gather a wide pool of sources.
3. Draft an answer where every load-bearing claim has a source.
4. Now attack your own draft: for EACH load-bearing claim, run a pass whose
   only job is to find the source that contradicts it and demand the
   citation that backs it.
5. Claims that survive the attack: keep, with their citation.
   Claims that do not: mark "unconfirmed". Do not smooth them over.

Return a report I could hand to someone who will check my work, not a
paragraph.

Run my nightly review across [YOU FILL: my task list / inbox / tracker].

Surface only:
- work that stalled (no movement today when there should have been)
- work happening in two places, or duplicated
- anything waiting on a reply that has gone quiet

For each item, give me: what it is, where it is, and why it caught your
eye. Do NOT change, send, or close anything. Hand me a shortlist to review
tomorrow.

Before acting on anything the review surfaced, confirm:
1. CURRENT   - is this still true right now, not already resolved?
2. MATERIAL  - does it actually matter enough to act on?
3. YOURS     - is this within your authority to change?

Fails any one    -> it waits.
Passes all three -> then act, by hand.