response-drafting-agent

Take any reply to a cold email (positive, question, objection, counter, referral, auto-reply, or dead-end) and, within 10-15 minutes, produce the right next move - a conversational reply, a research-grounded lead magnet PDF, both, or nothing. Offer-agnostic; recommends a path and waits for the operator to confirm before doing the heavy work. Use when the user pastes a prospect's reply to a cold email and wants the reply drafted.

SKILL.md

---
name: response-drafting-agent
description: Take any reply to a cold email (positive, question, objection, counter, referral, auto-reply, or dead-end) and, within 10-15 minutes, produce the right next move - a conversational reply, a research-grounded lead magnet PDF, both, or nothing. Offer-agnostic; recommends a path and waits for the operator to confirm before doing the heavy work. Use when the user pastes a prospect's reply to a cold email and wants the reply drafted.
metadata:
  title: Response Drafting Agent
  author: Christian Plascencia (Pipeline.tech)
  source: https://gamma.app/docs/AI-GTM-Masterclass-64rzstqd8tt14w9
  note: Shared in full in the free AI GTM Masterclass deck. Headings and list markers restored; the text is unchanged.
---

# Response Drafting Agent

## Purpose

Take any reply to a cold email (positive, question, objection, counter, referral, auto-reply, or dead-end) and, within 10-15 minutes, produce the right next move: a conversational reply in Shaw's voice, a research-grounded lead magnet, both, or nothing. The skill is offer-agnostic. The sender business, offer, and infra spec are supplied per run. Every asset is grounded in deep research on the prospect's company. No generic output.

The skill recommends a path, then waits for the operator to confirm before doing the heavy work.

## How to trigger this skill (read first)

The operator pastes the structured input block below. Fields marked optional can be left blank. If a required field is missing or marked [todo], ask for it before proceeding. Do not guess.

### Input template

```
SENDER BUSINESS URL: [https://yourdomain.com]   (the business sending the email)
OFFER ONE-LINER: [optional override; what you sell, one line]
WHAT THE EMAIL PROMISED: [e.g. "sample system: infra + accounts + copy" | "3 campaign ideas" | "nothing specific, just a call"]

INFRA SPEC (only if a system/infra magnet is in play):
  DAILY VOLUME TARGET: [default 1,000]
  SPLIT: [default 70% google / 30% microsoft]
  SEQUENCER: [EmailBison | PlusVibe | other]

PROSPECT EMAIL: [prefix@domain.com]
PROSPECT NAME: [First Last]
PROSPECT TITLE: [role]

ORIGINAL EMAIL BODY:
"""
[paste the body of the cold email you sent them]
"""

PROSPECT REPLY:
"""
[paste their full reply text]
"""
```

Required: SENDER BUSINESS URL, WHAT THE EMAIL PROMISED, PROSPECT EMAIL, ORIGINAL EMAIL BODY, PROSPECT REPLY. Name and title are strongly preferred but the run can proceed with title only if name is unknown.

If the operator pastes unstructured text, request the structured template before proceeding.

## Locked rules: do not override

### Voice rules (apply to ALL customer-facing output: reply email, copy templates, any text the prospect sees)

- all lowercase
- no em dashes ever; use commas, periods, or line breaks
- no emoji
- no hashtags
- short declarative sentences
- never start a hook with "i"
- peer-to-peer practitioner tone, not thought leader
- spaces between lines for readability
- roughly 280 character max per paragraph

SKILL.md instructions and internal process notes are not customer-facing and can use normal formatting.

### Reply email hard limits

- Under 100 words total (excluding signature).
- Propose 2-3 specific meeting times within the next 2-4 business days (afternoon preferred). Never "let me know what works." Compute days from today's date; do not hardcode.
- Must give a concrete reason why the call is valuable (e.g. "walk through the infra plan for a 1k/day setup"). Never "let's chat."
- Never re-pitch the offer. The cold email already did that. The reply delivers and proposes the call.
- Never repeat one-off references that lived only in the cold email (a P.S. hook, a named tool in the footer, etc.). Those stay in the cold email.

### Infrastructure math (parameterized, computed from INFRA SPEC)

Rates are fixed. Volume and split are inputs.

- Google layer: 15 emails/inbox/day, 2 inboxes per domain.
- Microsoft layer: 3 emails/account/day, 50 accounts per domain, 1 domain per Azure tenant.

Formula:

```
google_inboxes = round up to nearest even of (volume × google_split ÷ 15); google_domains = google_inboxes ÷ 2
ms_accounts = ceil(volume × ms_split ÷ 3); ms_domains = ceil(ms_accounts ÷ 50); ms_tenants = ms_domains
total_domains = google_domains + ms_domains; total_inboxes = google_inboxes + ms_accounts
```

Locked default check: at volume 1,000 and the default 70/30 split this resolves to the v2 baseline exactly: 24 google domains (48 inboxes), 2 MS domains (100 accounts), 2 Azure tenants, for a total of 26 domains, 148 inboxes. Sequencer: EmailBison or PlusVibe. Use these verbatim whenever volume is 1,000 at the default split.

### Lead magnet rules

- Every campaign idea, target account, persona, play, and piece of copy ties back to a specific, named signal pulled from research. No generic B2B fluff.
- Copy is written AS the prospect's company TO their prospects, not FROM the sender TO the prospect.
- Named humans must be LinkedIn-verifiable from research. If a name cannot be verified, use title only and mark the cell [to source via Clay]. Never fabricate.

## The gated flow (overview)

1. Parse and classify the reply.
2. Light recon, just enough to judge fit and pick a path.
3. Recommend a path with a one-line reason. If a magnet, name the shape and show the outline.
4. Wait for the operator to confirm.
5. On confirm: run deep research (only if a magnet was approved), build the asset(s), draft the reply.
6. Ultrareview, then deliver.

Heavy research and PDF building never run before step 4.

## Workflow

### Step 0 — Parse and classify reply (30 seconds)

1. Validate required fields. If missing, ask.
2. Extract {domain} from PROSPECT EMAIL (everything after @).
3. Classify the PROSPECT REPLY as one of seven types:
   - positive — "sure," "please send," "go for it," "interested"
   - question — asked about pricing, scope, timeline, fit, tools
   - objection — raised a concern, doubt, or constraint
   - counter-offer — proposed a different scope, format, or next step
   - referral / wrong person — "not me, talk to X," "I've forwarded this"
   - auto-reply / OOO — out of office, autoresponder, ticket bounce
   - negative / dead-end — "not interested," "unsubscribe," "stop," "remove me"

Output one confirmation line: `parsed. domain: {domain}. reply type: {type}. running recon.`

### Step 1 — Light recon (1-2 minutes)

Goal: enough context to recommend a path and pick a magnet shape. Not the full deep dive yet.

Tools: web_search and web_fetch only.

- web_fetch: SENDER BUSINESS URL once, to ground the offer and positioning (skip if OFFER ONE-LINER is supplied and clear).
- web_fetch: the prospect homepage (https://{domain}).
- 1-2 web_search queries to gauge stage, sales motion, and obvious recent signals.

Form a quick read: what they sell, who they sell to, whether the offer is a plausible fit, and which magnet shape would apply if the operator wants one.

#### Step 1 error handling

- Homepage errors or is sparse: try the LinkedIn company page or Crunchbase, else read Google snippets.
- Prospect is obscure with no web presence: ask the operator for a one-liner or LinkedIn URL before recommending a magnet.

### Step 2 — Recommend a path, then stop (30 seconds)

Map reply type to a recommended path:

| Reply type | Default recommendation |
| --- | --- |
| positive | reply + magnet (if the email promised an asset), else reply only |
| question | reply that answers it (+ magnet if it addresses the question or was promised) |
| objection | reply only, lower commitment. magnet only if it directly answers the concern |
| counter-offer | reply only, time aligned to their counter |
| referral / wrong person | short reply asking for the right contact. no magnet |
| auto-reply / OOO | log only. suggest a follow-up date after they return |
| negative / dead-end | output nothing. one log line, no reply, no asset |

Output to the operator (this is the gate):

```
reply type: {type}
recommendation: {reply only | reply + magnet | log only}
why: {one line}

[if magnet] magnet shape: {A outbound-system | B strategy}
[if magnet] outline:
  - {bullet 1}
  - {bullet 2}
  - {bullet 3}
  - {bullet 4}
  - {bullet 5}

confirm to proceed, or tell me what to change.
```

For negative / dead-end and auto-reply / OOO, do not gate. Just emit the log line and stop:

- dead-end: `dead-end reply from {name} at {domain}. no response sent. logged.`
- ooo: `auto-reply from {name} at {domain}. no action. suggest re-touch on {date after return}.`

Wait for confirmation before Step 3. If the operator changes the path or shape, honor it.

### Step 3 — On confirm: deep research + build

Run only after the operator confirms a path that includes an asset.

#### Step 3.1 — Deep research (ultrathink, 3-5 minutes; skip if reply-only was confirmed)

Minimum search sequence:

- web_search: {company name} case study
- web_search: {company name} hiring sdr OR {company name} head of sales
- web_search: {company name} funding news 2025 2026
- web_search: {company name} linkedin
- Targeted searches to close gaps (competitor comparisons, G2/Capterra, exec LinkedIn posts)

Coverage: what they sell, ICP, pricing/deal size if surfaced, stage, headcount and hiring signals, sales motion (founder-led / SDR / PLG / inbound / channel / hybrid), signs of a current outbound program, sales stack where inferable, core value prop in their words, named competitors, and anything from the last 90 days (weighted heaviest).

Ultrathink checkpoint (mandatory): write a 5-bullet internal summary. Can you name at least 3 distinct, specific signals that could each drive a campaign idea or play?

A signal is concrete: "hired a VP Sales from Gong in March who posts about multi-threaded outbound," or "dropped mid-market pricing in Q1 2026."

A signal is NOT an ICP label: "sells SaaS to B2B companies." If the 3-signal bar is not met, keep searching. If research honestly yields nothing specific after the full sequence, do not fabricate. Cap specificity to what is known and surface the gap in Step 6.

#### Step 3.2 — Build the asset

Pick the magnet shape from the offer and what the email promised. State the shape in Step 2; build it here.

**Shape A — Outbound system** (offer is infra / outbound build / "system, accounts, copy")

Build a PDF with these sections in order:

1. Summary (6-7 sentences). What the prospect sells, who they sell to, the outbound gap inferred from research, what this document outlines. Grounded in specific findings.
2. 10 target accounts. Real, named companies matching the prospect's ICP. Each row: company name, 1-sentence reason they fit (tie to a signal), rough size/stage.
3. 30 named humans (tier 1 / tier 2 / tier 3 across the 10 accounts): decision maker, decision influencer, champion. Each: name, title, company. LinkedIn-verifiable or [to source via Clay].
4. 3 copy templates (one per tier), written AS the prospect's company TO that tier. Each: subject line, 60-90 word body, voice rules, a trigger specific to that tier's day-to-day.
5. Infrastructure plan. Use the parameterized math from INFRA SPEC. Show daily volume, the google and microsoft layers with their domain/inbox counts, Azure tenant structure, totals, sequencer, and a short warmup note (2-3 weeks).

**Shape B — Strategy** (offer is advisory / GTM strategy / no infra build)

Build a PDF with these sections in order:

1. Summary (6-7 sentences). What the prospect sells, who they sell to, the strategic gap or opportunity inferred from research, what this document outlines.
2. Situation + gap analysis. 3-4 specific observations from research, each naming the signal it rests on.
3. 3-5 recommended plays. Each: play name (4-6 words), the signal that triggers it, target persona, the angle in 2-3 sentences.
4. One sample deliverable or framework. A worked example the prospect can react to (a sequence skeleton, a positioning angle, a sample subject + opener, or a one-page framework). Grounded, not generic.
5. What working together looks like. Short scope and first-30-days outline. No forced accounts/personas/infra math.

If the offer does not cleanly fit either shape, default to Shape B and flag it in Step 6.

PDF file name: `{company_slug}_{shapeA:pipeline_system|shapeB:gtm_strategy}_preview.pdf` Save to: `/mnt/user-data/outputs/`

Before writing any PDF, read `/mnt/skills/public/pdf/SKILL.md`.

### Step 4 — Draft the reply email (1 minute)

Structure for all reply types:

1. Opener that acknowledges their reply naturally (no corporate "thanks for getting back to me").
2. Deliver the asset (or, if reply-only, move straight to the CTA).
3. Concrete CTA with 2-3 specific afternoon times.
4. Signoff with %signature%.

Reply type branches:

- positive → opener acknowledges in one line. deliver asset. propose call.
- question → opener answers the question in 1-2 sentences. deliver asset if relevant. propose call.
- objection → opener acknowledges the concern without defending (one line). deliver asset only if it addresses the concern. lower-commitment CTA ("open to a 15-min walk-through if useful?").
- counter-offer → opener accepts or redirects their counter in one line. deliver asset. propose a time aligned to their counter.
- referral / wrong person → opener thanks briefly and asks for the right person by name or title. no asset. no hard CTA, just the ask.

Asset delivery line (when a magnet was built): one line like `put together a sample system for {company}, attached here.`

CTA examples (in voice rules):

- weak (never use): "let me know what works," "happy to chat," "let me know if this is interesting"
- strong: "open for a 15-min walk-through tuesday or wednesday afternoon?" / "worth a quick call to walk through the infra plan. thursday 2pm or friday 3pm?" / "free for 20 min next week to go through the account list. monday or tuesday afternoon?"

Pick 2-3 afternoon slots within the next 2-4 business days, computed from today's date.

Hard caps: under 100 words, no paragraph over 280 characters, no re-pitch.

### Step 5 — Ultrareview (mandatory before output, 1 minute)

If any item fails, fix before delivering.

Research depth (if a magnet was built)

- 3 specific signals named from research, not ICP labels
- sender offer grounded + prospect homepage + supporting sources visited
- ICP understood well enough to pick 10 real accounts (Shape A) or 3-5 grounded plays (Shape B)

Lead magnet quality

- every idea / play / copy template references a specific signal
- Shape A: 10 target accounts are real named companies; 30 personas real where verifiable, flagged otherwise; infra math matches the formula
- copy written AS the prospect's company, not as the sender
- correct shape for the offer

Voice

- all customer-facing text lowercase
- zero em dashes, emoji, hashtags
- no hook starts with "i"

Reply email

- under 100 words
- 2-3 specific afternoon times proposed
- concrete reason the call is valuable
- reply type structure honored
- no re-pitch, no cold-email-only references repeated

### Step 6 — Output to operator (no narration)

Deliver in this order:

1. Reply email body in a code block for copy-paste (omit if path was log only or referral with no asset still includes the reply).
2. If a PDF was built: the file via present_files.
3. One-line summary plus open items:
   - personas flagged [to source via Clay] (count)
   - research gaps surfaced (if any)
   - magnet shape used, if it was a fallback to B
   - suggested follow-up date (3 business days from today) if no response

Do not narrate research or process. Just deliver.

### Step 7 — Follow-up logic (for operator's awareness)

If the prospect does not reply within 3 business days, send one check-in from the same inbox, 50 words max:

```
{first_name}, circling back on the {asset type} i sent. still worth a quick walk-through?

{day} or {day} afternoon work?

%signature%
```

One follow-up only. If still no reply, move on. For OOO replies, set the check-in for after their return date instead.

## Time budget per run

- Step 0 (parse): 30 sec
- Step 1 (recon): 1-2 min
- Step 2 (recommend + gate): 30 sec, then wait
- Step 3 (deep research + build): 5-7 min for a magnet, near zero for reply-only
- Step 4 (reply): 1 min
- Step 5 (ultrareview): 1 min
- Step 6 (deliver): 30 sec

Target: ~5 min for reply-only, 12-15 min for a magnet. Quality over speed. Do not rush the ultrareview.

## Common mistakes to avoid

Recon / research

- Doing the full deep dive before the operator confirms a path. Recon is light; deep research waits for the gate.
- Treating ICP labels as signals. Signals are concrete events or states.
- Skipping the prospect homepage and jumping to search.

Asset

- Fabricating named humans. Mark [to source via Clay] instead.
- Recalculating infra rates. Rates are fixed; only volume and split vary.
- Writing copy FROM the sender TO the prospect. Copy is AS the prospect TO their prospects.
- Forcing Shape A structure onto a strategy offer. Pick the shape from the offer.

Reply

- Vague meeting proposals. Always 2-3 specific afternoon slots.
- Exceeding 100 words. Trim to the highest-value 2 sentences.
- Re-pitching the offer or repeating cold-email-only references.
- Starting a hook with "i."

Process

- Narrating research or process in the output. Just deliver.
- Skipping the gate or the ultrareview.
- Hardcoding meeting day names. Compute from today.
- Sending anything on a dead-end reply. Log only.

1 file

SKILL.md17.1 KB

--- name: response-drafting-agent description: Take any reply to a cold email (positive, question, objection, counter, referral, auto-reply, or dead-end) and, within 10-15 minutes, produce the right next move - a conversational reply, a research-grounded lead magnet PDF, both, or nothing. Offer-agnostic; recommends a path and waits for the operator to confirm before doing the heavy work. Use when the user pastes a prospect's reply to a cold email and wants the reply drafted. metadata: title: Response Drafting Agent author: Christian Plascencia (Pipeline.tech) source: https://gamma.app/docs/AI-GTM-Masterclass-64rzstqd8tt14w9 note: Shared in full in the free AI GTM Masterclass deck. Headings and list markers restored; the text is unchanged. --- # Response Drafting Agent ## Purpose Take any reply to a cold email (positive, question, objection, counter, referral, auto-reply, or dead-end) and, within 10-15 minutes, produce the right next move: a conversational reply in Shaw's voice, a research-grounded lead magnet, both, or nothing. The skill is offer-agnostic. The sender business, offer, and infra spec are supplied per run. Every asset is grounded in deep research on the prospect's company. No generic output. The skill recommends a path, then waits for the operator to confirm before doing the heavy work. ## How to trigger this skill (read first) The operator pastes the structured input block below. Fields marked optional can be left blank. If a required field is missing or marked [todo], ask for it before proceeding. Do not guess. ### Input template ``` SENDER BUSINESS URL: [https://yourdomain.com] (the business sending the email) OFFER ONE-LINER: [optional override; what you sell, one line] WHAT THE EMAIL PROMISED: [e.g. "sample system: infra + accounts + copy" | "3 campaign ideas" | "nothing specific, just a call"] INFRA SPEC (only if a system/infra magnet is in play): DAILY VOLUME TARGET: [default 1,000] SPLIT: [default 70% google / 30% microsoft] SEQUENCER: [EmailBison | PlusVibe | other] PROSPECT EMAIL: [prefix@domain.com] PROSPECT NAME: [First Last] PROSPECT TITLE: [role] ORIGINAL EMAIL BODY: """ [paste the body of the cold email you sent them] """ PROSPECT REPLY: """ [paste their full reply text] """ ``` Required: SENDER BUSINESS URL, WHAT THE EMAIL PROMISED, PROSPECT EMAIL, ORIGINAL EMAIL BODY, PROSPECT REPLY. Name and title are strongly preferred but the run can proceed with title only if name is unknown. If the operator pastes unstructured text, request the structured template before proceeding. ## Locked rules: do not override ### Voice rules (apply to ALL customer-facing output: reply email, copy templates, any text the prospect sees) - all lowercase - no em dashes ever; use commas, periods, or line breaks - no emoji - no hashtags - short declarative sentences - never start a hook with "i" - peer-to-peer practitioner tone, not thought leader - spaces between lines for readability - roughly 280 character max per paragraph SKILL.md instructions and internal process notes are not customer-facing and can use normal formatting. ### Reply email hard limits - Under 100 words total (excluding signature). - Propose 2-3 specific meeting times within the next 2-4 business days (afternoon preferred). Never "let me know what works." Compute days from today's date; do not hardcode. - Must give a concrete reason why the call is valuable (e.g. "walk through the infra plan for a 1k/day setup"). Never "let's chat." - Never re-pitch the offer. The cold email already did that. The reply delivers and proposes the call. - Never repeat one-off references that lived only in the cold email (a P.S. hook, a named tool in the footer, etc.). Those stay in the cold email. ### Infrastructure math (parameterized, computed from INFRA SPEC) Rates are fixed. Volume and split are inputs. - Google layer: 15 emails/inbox/day, 2 inboxes per domain. - Microsoft layer: 3 emails/account/day, 50 accounts per domain, 1 domain per Azure tenant. Formula: ``` google_inboxes = round up to nearest even of (volume × google_split ÷ 15); google_domains = google_inboxes ÷ 2 ms_accounts = ceil(volume × ms_split ÷ 3); ms_domains = ceil(ms_accounts ÷ 50); ms_tenants = ms_domains total_domains = google_domains + ms_domains; total_inboxes = google_inboxes + ms_accounts ``` Locked default check: at volume 1,000 and the default 70/30 split this resolves to the v2 baseline exactly: 24 google domains (48 inboxes), 2 MS domains (100 accounts), 2 Azure tenants, for a total of 26 domains, 148 inboxes. Sequencer: EmailBison or PlusVibe. Use these verbatim whenever volume is 1,000 at the default split. ### Lead magnet rules - Every campaign idea, target account, persona, play, and piece of copy ties back to a specific, named signal pulled from research. No generic B2B fluff. - Copy is written AS the prospect's company TO their prospects, not FROM the sender TO the prospect. - Named humans must be LinkedIn-verifiable from research. If a name cannot be verified, use title only and mark the cell [to source via Clay]. Never fabricate. ## The gated flow (overview) 1. Parse and classify the reply. 2. Light recon, just enough to judge fit and pick a path. 3. Recommend a path with a one-line reason. If a magnet, name the shape and show the outline. 4. Wait for the operator to confirm. 5. On confirm: run deep research (only if a magnet was approved), build the asset(s), draft the reply. 6. Ultrareview, then deliver. Heavy research and PDF building never run before step 4. ## Workflow ### Step 0 — Parse and classify reply (30 seconds) 1. Validate required fields. If missing, ask. 2. Extract {domain} from PROSPECT EMAIL (everything after @). 3. Classify the PROSPECT REPLY as one of seven types: - positive — "sure," "please send," "go for it," "interested" - question — asked about pricing, scope, timeline, fit, tools - objection — raised a concern, doubt, or constraint - counter-offer — proposed a different scope, format, or next step - referral / wrong person — "not me, talk to X," "I've forwarded this" - auto-reply / OOO — out of office, autoresponder, ticket bounce - negative / dead-end — "not interested," "unsubscribe," "stop," "remove me" Output one confirmation line: `parsed. domain: {domain}. reply type: {type}. running recon.` ### Step 1 — Light recon (1-2 minutes) Goal: enough context to recommend a path and pick a magnet shape. Not the full deep dive yet. Tools: web_search and web_fetch only. - web_fetch: SENDER BUSINESS URL once, to ground the offer and positioning (skip if OFFER ONE-LINER is supplied and clear). - web_fetch: the prospect homepage (https://{domain}). - 1-2 web_search queries to gauge stage, sales motion, and obvious recent signals. Form a quick read: what they sell, who they sell to, whether the offer is a plausible fit, and which magnet shape would apply if the operator wants one. #### Step 1 error handling - Homepage errors or is sparse: try the LinkedIn company page or Crunchbase, else read Google snippets. - Prospect is obscure with no web presence: ask the operator for a one-liner or LinkedIn URL before recommending a magnet. ### Step 2 — Recommend a path, then stop (30 seconds) Map reply type to a recommended path: | Reply type | Default recommendation | | --- | --- | | positive | reply + magnet (if the email promised an asset), else reply only | | question | reply that answers it (+ magnet if it addresses the question or was promised) | | objection | reply only, lower commitment. magnet only if it directly answers the concern | | counter-offer | reply only, time aligned to their counter | | referral / wrong person | short reply asking for the right contact. no magnet | | auto-reply / OOO | log only. suggest a follow-up date after they return | | negative / dead-end | output nothing. one log line, no reply, no asset | Output to the operator (this is the gate): ``` reply type: {type} recommendation: {reply only | reply + magnet | log only} why: {one line} [if magnet] magnet shape: {A outbound-system | B strategy} [if magnet] outline: - {bullet 1} - {bullet 2} - {bullet 3} - {bullet 4} - {bullet 5} confirm to proceed, or tell me what to change. ``` For negative / dead-end and auto-reply / OOO, do not gate. Just emit the log line and stop: - dead-end: `dead-end reply from {name} at {domain}. no response sent. logged.` - ooo: `auto-reply from {name} at {domain}. no action. suggest re-touch on {date after return}.` Wait for confirmation before Step 3. If the operator changes the path or shape, honor it. ### Step 3 — On confirm: deep research + build Run only after the operator confirms a path that includes an asset. #### Step 3.1 — Deep research (ultrathink, 3-5 minutes; skip if reply-only was confirmed) Minimum search sequence: - web_search: {company name} case study - web_search: {company name} hiring sdr OR {company name} head of sales - web_search: {company name} funding news 2025 2026 - web_search: {company name} linkedin - Targeted searches to close gaps (competitor comparisons, G2/Capterra, exec LinkedIn posts) Coverage: what they sell, ICP, pricing/deal size if surfaced, stage, headcount and hiring signals, sales motion (founder-led / SDR / PLG / inbound / channel / hybrid), signs of a current outbound program, sales stack where inferable, core value prop in their words, named competitors, and anything from the last 90 days (weighted heaviest). Ultrathink checkpoint (mandatory): write a 5-bullet internal summary. Can you name at least 3 distinct, specific signals that could each drive a campaign idea or play? A signal is concrete: "hired a VP Sales from Gong in March who posts about multi-threaded outbound," or "dropped mid-market pricing in Q1 2026." A signal is NOT an ICP label: "sells SaaS to B2B companies." If the 3-signal bar is not met, keep searching. If research honestly yields nothing specific after the full sequence, do not fabricate. Cap specificity to what is known and surface the gap in Step 6. #### Step 3.2 — Build the asset Pick the magnet shape from the offer and what the email promised. State the shape in Step 2; build it here. **Shape A — Outbound system** (offer is infra / outbound build / "system, accounts, copy") Build a PDF with these sections in order: 1. Summary (6-7 sentences). What the prospect sells, who they sell to, the outbound gap inferred from research, what this document outlines. Grounded in specific findings. 2. 10 target accounts. Real, named companies matching the prospect's ICP. Each row: company name, 1-sentence reason they fit (tie to a signal), rough size/stage. 3. 30 named humans (tier 1 / tier 2 / tier 3 across the 10 accounts): decision maker, decision influencer, champion. Each: name, title, company. LinkedIn-verifiable or [to source via Clay]. 4. 3 copy templates (one per tier), written AS the prospect's company TO that tier. Each: subject line, 60-90 word body, voice rules, a trigger specific to that tier's day-to-day. 5. Infrastructure plan. Use the parameterized math from INFRA SPEC. Show daily volume, the google and microsoft layers with their domain/inbox counts, Azure tenant structure, totals, sequencer, and a short warmup note (2-3 weeks). **Shape B — Strategy** (offer is advisory / GTM strategy / no infra build) Build a PDF with these sections in order: 1. Summary (6-7 sentences). What the prospect sells, who they sell to, the strategic gap or opportunity inferred from research, what this document outlines. 2. Situation + gap analysis. 3-4 specific observations from research, each naming the signal it rests on. 3. 3-5 recommended plays. Each: play name (4-6 words), the signal that triggers it, target persona, the angle in 2-3 sentences. 4. One sample deliverable or framework. A worked example the prospect can react to (a sequence skeleton, a positioning angle, a sample subject + opener, or a one-page framework). Grounded, not generic. 5. What working together looks like. Short scope and first-30-days outline. No forced accounts/personas/infra math. If the offer does not cleanly fit either shape, default to Shape B and flag it in Step 6. PDF file name: `{company_slug}_{shapeA:pipeline_system|shapeB:gtm_strategy}_preview.pdf` Save to: `/mnt/user-data/outputs/` Before writing any PDF, read `/mnt/skills/public/pdf/SKILL.md`. ### Step 4 — Draft the reply email (1 minute) Structure for all reply types: 1. Opener that acknowledges their reply naturally (no corporate "thanks for getting back to me"). 2. Deliver the asset (or, if reply-only, move straight to the CTA). 3. Concrete CTA with 2-3 specific afternoon times. 4. Signoff with %signature%. Reply type branches: - positive → opener acknowledges in one line. deliver asset. propose call. - question → opener answers the question in 1-2 sentences. deliver asset if relevant. propose call. - objection → opener acknowledges the concern without defending (one line). deliver asset only if it addresses the concern. lower-commitment CTA ("open to a 15-min walk-through if useful?"). - counter-offer → opener accepts or redirects their counter in one line. deliver asset. propose a time aligned to their counter. - referral / wrong person → opener thanks briefly and asks for the right person by name or title. no asset. no hard CTA, just the ask. Asset delivery line (when a magnet was built): one line like `put together a sample system for {company}, attached here.` CTA examples (in voice rules): - weak (never use): "let me know what works," "happy to chat," "let me know if this is interesting" - strong: "open for a 15-min walk-through tuesday or wednesday afternoon?" / "worth a quick call to walk through the infra plan. thursday 2pm or friday 3pm?" / "free for 20 min next week to go through the account list. monday or tuesday afternoon?" Pick 2-3 afternoon slots within the next 2-4 business days, computed from today's date. Hard caps: under 100 words, no paragraph over 280 characters, no re-pitch. ### Step 5 — Ultrareview (mandatory before output, 1 minute) If any item fails, fix before delivering. Research depth (if a magnet was built) - 3 specific signals named from research, not ICP labels - sender offer grounded + prospect homepage + supporting sources visited - ICP understood well enough to pick 10 real accounts (Shape A) or 3-5 grounded plays (Shape B) Lead magnet quality - every idea / play / copy template references a specific signal - Shape A: 10 target accounts are real named companies; 30 personas real where verifiable, flagged otherwise; infra math matches the formula - copy written AS the prospect's company, not as the sender - correct shape for the offer Voice - all customer-facing text lowercase - zero em dashes, emoji, hashtags - no hook starts with "i" Reply email - under 100 words - 2-3 specific afternoon times proposed - concrete reason the call is valuable - reply type structure honored - no re-pitch, no cold-email-only references repeated ### Step 6 — Output to operator (no narration) Deliver in this order: 1. Reply email body in a code block for copy-paste (omit if path was log only or referral with no asset still includes the reply). 2. If a PDF was built: the file via present_files. 3. One-line summary plus open items: - personas flagged [to source via Clay] (count) - research gaps surfaced (if any) - magnet shape used, if it was a fallback to B - suggested follow-up date (3 business days from today) if no response Do not narrate research or process. Just deliver. ### Step 7 — Follow-up logic (for operator's awareness) If the prospect does not reply within 3 business days, send one check-in from the same inbox, 50 words max: ``` {first_name}, circling back on the {asset type} i sent. still worth a quick walk-through? {day} or {day} afternoon work? %signature% ``` One follow-up only. If still no reply, move on. For OOO replies, set the check-in for after their return date instead. ## Time budget per run - Step 0 (parse): 30 sec - Step 1 (recon): 1-2 min - Step 2 (recommend + gate): 30 sec, then wait - Step 3 (deep research + build): 5-7 min for a magnet, near zero for reply-only - Step 4 (reply): 1 min - Step 5 (ultrareview): 1 min - Step 6 (deliver): 30 sec Target: ~5 min for reply-only, 12-15 min for a magnet. Quality over speed. Do not rush the ultrareview. ## Common mistakes to avoid Recon / research - Doing the full deep dive before the operator confirms a path. Recon is light; deep research waits for the gate. - Treating ICP labels as signals. Signals are concrete events or states. - Skipping the prospect homepage and jumping to search. Asset - Fabricating named humans. Mark [to source via Clay] instead. - Recalculating infra rates. Rates are fixed; only volume and split vary. - Writing copy FROM the sender TO the prospect. Copy is AS the prospect TO their prospects. - Forcing Shape A structure onto a strategy offer. Pick the shape from the offer. Reply - Vague meeting proposals. Always 2-3 specific afternoon slots. - Exceeding 100 words. Trim to the highest-value 2 sentences. - Re-pitching the offer or repeating cold-email-only references. - Starting a hook with "i." Process - Narrating research or process in the output. Just deliver. - Skipping the gate or the ultrareview. - Hardcoding meeting day names. Compute from today. - Sending anything on a dead-end reply. Log only.