getaero-io/gtm-eng-skills / skills/deepline-gtm
deepline-gtm
UnverifiedGTM leads, enrichment, outreach, Plays, developer signals/Reo.dev. Providers: adyntel,affinity,ai_ark,akta,allegrow,amplemarket,apify,attention,attio,aviato,bettercontact,bigquery,bluesky,bounceban,brave,browserbase,builtwith,buyercaddy,cbinsights,clay,clickhouse,close,cloudflare,contactout,contextdev,crunchbase,crustdata-v3,customer_db,databricks,dataforseo,datagma,dealroom,deepline_ip_to_company,deepline_native,deeplineagent,deepseek,discolike,dnb,edges,emailbison,emailguard,enformion,enigma,exa,findymail,firecrawl,fireflies,firmable,forager,fullenrich,fundable,generic_http,getleads,gong,google_ads_audiences,google_workspace,govfiles,grain,hackernews,harvestapi,heyreach,hubspot,hunter,icypeas,instantly,intercom,ipqs,kaspr,kernel,kickbox,leadmagic,lemlist,limadata,linkedin_ads_audiences,luma,lusha,meerapi,meta_audiences,millionverifier,nooks,notion,openmart,opensosdata,openwebninja,outreach,parallel,particle,peopledatalabs,pitchbook,podscan,polymarket,postgres,predictleads,prospeo,quickenrich,rapidapi,….
writing-outreach.md
View on GitHub# Writing Outreach Skill
Use this skill when the task involves cold emails, personalization, lead scoring, qualification, sequence design, campaign copy, or inspection of enrichment results.
## What This Skill Does
- Loads ICP/criteria context from a local file (for example `./icp.md`) using the agent’s local file reader.
- Produces strict, structured qualification output:
- numeric score (`0-10`)
- fit label (for example `Strong fit: 8/10`)
- rationale summary
- question-by-question yes/no/unknown answers with confidence and rationale
- Produces strict, structured 4-step outbound email sequence:
- subject
- core value prop
- message body
- sequence-level rationale
## Required Inputs
- `prospect_payload`: Person + company context (JSON string/object in row)
- `icp.md` (or equivalent): Local qualification criteria and positioning constraints
## General Setup (Recommended)
Create a small context pack in your working directory before running this flow:
- `icp.md`: Ideal customer profile and disqualifiers
- `qualification_questions.md`: The exact questions you want scored
- `product_context.md`: Value props, proof points, and constraints
- `copy_rules.md`: Tone, banned claims, length limits, CTA style
Minimal folder layout:
```text
./context/
icp.md
qualification_questions.md
product_context.md
copy_rules.md
```
Example `icp.md` starter:
```markdown
# ICP
## Best-fit companies
- B2B SaaS, 200-5000 employees, multi-product GTM team
- Uses modern data stack and CRM-based workflows
## Best-fit personas
- VP/Head/Director in Marketing, RevOps, Sales Ops, GTM Ops
- Owns pipeline quality, segmentation, scoring, or campaign orchestration
## Core pains
- Slow GTM iteration due to analyst/engineering dependency
- Low trust in black-box scoring
- Weak signal-to-action workflow for reps
## Disqualifiers
- <20 employees
- Pure B2C motion
- No clear sales/marketing operations function
```
## Best Practices
- Keep `icp.md` specific and opinionated. Broad ICPs create generic copy.
- Separate facts vs assumptions in your context docs.
- Put proof points in `product_context.md` so emails can reference them cleanly.
- Define hard copy constraints in `copy_rules.md`:
tone, max words, banned phrases, allowed CTA types.
- Use `Unknown` for missing evidence in qualification instead of guessing.
- Run the QA prompt template before finalizing sequence copy.
- Version your context docs with the campaign (`context-q1-enterprise.md`, etc.).
## Output Contracts
### Qualification JSON Contract
```json
{
"data": {
"score": 8,
"score_label": "Strong fit: 8/10",
"fit_band": "STRONG_FIT",
"rationale": "Short summary of fit and caveats.",
"qualification": {
"answers": [
{
"question": "string",
"answer": "Yes",
"confidence": "HIGH",
"rationale": "string"
}
],
"summary": {
"positives": ["string"],
"risks": ["string"],
"next_checks": ["string"]
}
}
}
}
```
### Email Sequence JSON Contract
```json
{
"data": {
"emails": [
{
"step": 1,
"subject": "string",
"coreValueProp": "string",
"email": "string"
}
],
"sequence_rationale": "string"
}
}
```
## Personalization: native callbacks and AI
When you already have contact + company context in CSV columns, use a native Play callback for email generation when a deterministic template is enough. Use `deeplineagent` when you want AI help with reasoning, scoring, copy quality, or research.
**Critical: avoid mail-merge output.** If every email has the same structure with only `{{first_name}}` and `{{company_name}}` swapped, it's a template — not personalized outreach. Each email must reference something specific to the company (product, use case, industry, recent news). Use `company_description`, `one_liner`, `company_research`, or other enrichment columns in your JS template or `deeplineagent` prompt so each email is substantively different.
Author the generation column(s) as a custom play over your enriched CSV
(see `deepline plays --help`) with
one `withColumn` per output:
- Fast path: a callback-based `withColumn` applying a deterministic template to typed row values.
- AI path: a `deeplineagent` column with the research columns interpolated into
the prompt and a `jsonSchema` of `{subject, email}`.
- Research-first path: a research `deeplineagent` column, then the generation
column consuming it.
Notes:
- `{{company_research}}` is the safe form for AI prompts when the prior column came from `deeplineagent`.
- If you need a single field from prior research for deterministic template logic, extract it into its own scalar column with a native callback first, then reference that scalar column in later steps.
- Inside a callback, use `row["company_research"]` or `row.company_research`. Read typed row values directly instead of brace interpolation.
## Recommended Workflow
Write the two columns as a play, `qualify.play.ts`. A play reads only its
input, so the ICP and product context arrive as input fields:
```ts
import deepline from 'deepline';
// The output contracts above, as JSON Schemas.
const strictObject = (properties: Record<string, object>) => ({
type: 'object',
properties,
required: Object.keys(properties),
additionalProperties: false,
});
const QUALIFICATION_SCHEMA = strictObject({
data: strictObject({
score: { type: 'number' },
score_label: { type: 'string' },
fit_band: { type: 'string' },
rationale: { type: 'string' },
}),
});
const SEQUENCE_SCHEMA = strictObject({
data: strictObject({
emails: {
type: 'array',
items: strictObject({
step: { type: 'number' },
subject: { type: 'string' },
coreValueProp: { type: 'string' },
email: { type: 'string' },
}),
},
sequence_rationale: { type: 'string' },
}),
});
export const config = {
name: 'qualify-and-sequence',
description:
'Qualify prospects against the ICP and write a 4-step email sequence.',
};
export default async function (input: {
csv: string;
icp: string;
product: string;
}) {
const prospects = await deepline
.dataset('prospects', await deepline.csv(input.csv))
.withColumn('qualification_output', async (row) => {
const result = await deepline.tools.execute({
id: 'qualification',
tool: 'deeplineagent',
input: {
model: 'openai/gpt-5.6-luna',
prompt: `You are a B2B qualification analyst. Use only the provided evidence.
ICP context:
${input.icp}
Prospect payload:
${row.prospect_payload}`,
jsonSchema: QUALIFICATION_SCHEMA,
},
description: 'Score this prospect against the ICP.',
});
return result.toolResponse.rawV2.result.object;
})
.withColumn('email_sequence_output', async (row) => {
const result = await deepline.tools.execute({
id: 'email_sequence',
tool: 'deeplineagent',
input: {
model: 'openai/gpt-5.6-luna',
prompt: `You are a B2B email strategist. Write 4 concise emails tied to the qualification output and product context.
Product context:
${input.product}
Qualification payload:
${JSON.stringify(row.qualification_output)}
Prospect payload:
${row.prospect_payload}`,
jsonSchema: SEQUENCE_SCHEMA,
},
description: 'Write a 4-step email sequence for this prospect.',
});
return result.toolResponse.rawV2.result.object;
})
.run({ key: 'id' });
return { prospects };
}
```
Seed one prospect, then check and run the play:
```bash
deepline describe deeplineagent
cat > ./qualification_email_seed.csv <<'CSV'
id,prospect_payload
cloudera-rfoster,"{""person"":{""firstName"":""Rachael"",""lastName"":""Foster"",""title"":""Vice President AMER Field Marketing, Public Sector, Services, & Community""},""company"":{""name"":""Cloudera"",""domain"":""cloudera.com""}}"
CSV
jq -n --rawfile icp ./context/icp.md --rawfile product ./context/product_context.md \
'{csv: "qualification_email_seed.csv", icp: $icp, product: $product}' > ./qualify-input.json
deepline plays check qualify.play.ts
deepline run qualify.play.ts --input-file ./qualify-input.json \
--csv-out "qualified-$(date +%Y%m%d-%H%M%S).csv"
```
## Prompt Templates (Copy/Paste)
Use these as `prompt` values in `deeplineagent` input. Pair each prompt with a `jsonSchema` that matches the output contracts above. Recommended starting model: `openai/gpt-5.6-luna`.
### 1) Score + Fit Label Template
```json
{
"prompt": "You are a B2B scoring analyst.\nReturn strict JSON only:\n{\"data\":{\"score\":0,\"score_label\":\"\",\"fit_band\":\"\",\"scoring\":{\"weights\":[{\"factor\":\"\",\"weight\":0,\"evidence\":\"\",\"impact\":\"positive|negative|neutral\"}],\"confidence\":\"HIGH|MEDIUM|LOW\"},\"rationale\":\"\"}}\nRules:\n- score is integer 0-10\n- score_label format: \"Strong fit: X/10\" or \"Possible fit: X/10\" or \"Weak fit: X/10\"\n- fit_band one of: STRONG_FIT, POSSIBLE_FIT, WEAK_FIT\n- use evidence from provided context only; do not invent facts\n- default to higher recall unless strict matching is explicitly requested\nInputs:\nICP context: {{icp_context}}\nProspect payload: {{prospect_payload}}",
"model": "openai/gpt-5.6-luna"
}
```
### 2) Qualification QA + Rationale Template
```json
{
"prompt": "You are an ICP qualification analyst.\nReturn strict JSON only:\n{\"data\":{\"qualification\":{\"answers\":[{\"question\":\"\",\"answer\":\"Yes|No|Unknown\",\"confidence\":\"HIGH|MEDIUM|LOW\",\"rationale\":\"\"}],\"summary\":{\"positives\":[\"\"],\"risks\":[\"\"],\"next_checks\":[\"\"]}},\"rationale\":\"\"}}\nRules:\n- keep answers short and explicit\n- each rationale must reference concrete evidence from context\n- mark Unknown when evidence is missing\n- avoid strict/exact matching unless explicitly asked\nInputs:\nICP questions: {{qualification_questions}}\nICP context: {{icp_context}}\nProspect payload: {{prospect_payload}}",
"model": "openai/gpt-5.6-luna"
}
```
### 3) 4-Step Email Sequence Design Template
```json
{
"prompt": "You are a B2B email strategist.\nReturn strict JSON only:\n{\"data\":{\"emails\":[{\"step\":1,\"subject\":\"\",\"coreValueProp\":\"\",\"email\":\"\"}],\"sequence_rationale\":\"\"}}\nRules:\n- exactly 4 emails with step 1..4\n- each email must map to a pain or risk from qualification summary\n- concise style, no fluff, no markdown\n- personalization must reference role/company context from prospect payload\n- avoid claims not supported by inputs\nInputs:\nProduct context: {{product_context}}\nQualification output: {{qualification_output}}\nProspect payload: {{prospect_payload}}",
"model": "openai/gpt-5.6-luna"
}
```
### 4) Subject Line Variant Generator Template
```json
{
"prompt": "You are a B2B subject line writer.\nReturn strict JSON only:\n{\"data\":{\"subjects\":[{\"variant\":\"\",\"angle\":\"pain|outcome|proof|curiosity\",\"why_it_matches\":\"\"}]}}\nRules:\n- create 8 variants max 7 words each\n- no clickbait, no ALL CAPS, no exclamation marks\n- tie each variant to qualification rationale\nInputs:\nQualification output: {{qualification_output}}\nProspect payload: {{prospect_payload}}",
"model": "openai/gpt-5.6-luna"
}
```
### 5) Email Quality Critique Template (Optional QA Pass)
```json
{
"prompt": "You are a cold email QA editor.\nReturn strict JSON only:\n{\"data\":{\"issues\":[{\"step\":1,\"severity\":\"HIGH|MEDIUM|LOW\",\"issue\":\"\",\"fix\":\"\"}],\"revised_emails\":[{\"step\":1,\"subject\":\"\",\"coreValueProp\":\"\",\"email\":\"\"}]}}\nRules:\n- flag vague claims, unsupported assertions, and weak personalization\n- keep original structure and tighten only where needed\nInputs:\nQualification output: {{qualification_output}}\nEmail sequence output: {{email_sequence_output}}\nProspect payload: {{prospect_payload}}",
"model": "openai/gpt-5.6-luna"
}
```
## Guardrails
- Keep qualification deterministic and evidence-based from provided context.
- Prefer high recall unless strict matching is explicitly requested.
- Keep outputs strict JSON (no markdown wrappers).
- Keep email copy concise, specific to role/company context, and grounded in qualification rationale.
## CSV Inspection
Use these commands to interact with `deepline csv` directly for inspecting and debugging enrichment results.
### Open an existing CSV
```bash
deepline csv render --csv leads.csv --open
```
- Use `--open` to launch the CSV viewer.
### Inspect rows (`deepline csv show`)
Keep the complete export, including intermediate evidence; select columns for
display with `csv show` instead of changing the exported file. See
`deepline csv show -h` for options.
- `--format json|table|csv` applies to both row display and summary; JSON is default.
- `--columns a,b` includes named columns; `--exclude-columns c` omits them.
Exclusions win. Unknown names or an empty selection fail; the file is unchanged.
- `--verbose` shows full cell values in table output, not additional columns.
- `--summary` returns one row per selected column with present/missing counts,
present percentage, unique count, top repeated values and other count. It
analyzes all file rows by default; presence is not verification or a miss reason.
- `--rows start:end` is inclusive. Normal display defaults to `0:19`; an
explicit range also limits the summary.
```bash
deepline csv show leads.csv
deepline csv show leads.csv --columns first_name,email --format table --rows 0:10
deepline csv show leads.csv --summary
```
Use actual column names from the file. CSV JSON display still contains string
cells. For native nested objects and arrays, use `runs export --format json`
as described in `deepline runs export --help`.
### Re-run a column
`deepline csv` is local inspection only. Inspect the existing export and run
before proposing recomputation. A new run or forced refresh needs execution
authority and may spend again; completed receipt reuse is not a free-rerun
guarantee. See `deepline run --help` and `deepline runs rerun --help`.
### CLI-only debug posture
- These commands inspect local files; they run nothing. Deterministic
processing is allowed; preserve the original and label derived outputs.
- To add columns or providers, use a prebuilt play from `deepline search` or a custom play.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| deprecation-and-migrationaddyosmani/agent-skillsManages deprecation and migration. Use when removing old systems, APIs, or features. Use when migrating users from one implementation to another. Use when migrating a database schema in production, such as renaming or dropping a column without downtime (expand/contract). Use when deciding whether to maintain or sunset existing code. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| performance-optimizationaddyosmani/agent-skillsOptimizes application performance across frontend, backend, queries, and databases. Use when performance requirements exist, when you suspect performance regressions, when Core Web Vitals or load times need improvement, when N+1 query patterns need fixing, or when profiling reveals bottlenecks. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| agent-cost-reportthedotmack/claude-memBelievable agent cost report for any period, default the last 7 full days PT, not counting today. Measured tokens from Claude Code transcripts priced at OpenRouter list prices (ESTIMATED), measured provider spend when a sanctioned source exists, note-taker cost separate, Timing-style HTML/PDF plus report.json, line-items.csv, evidence.json. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| agent-cost-reportthedotmack/claude-memBelievable agent cost report for any period, default the last 7 full days PT, not counting today. Measured tokens from Claude Code transcripts priced at OpenRouter list prices (ESTIMATED), measured provider spend when a sanctioned source exists, note-taker cost separate, Timing-style HTML/PDF plus report.json, line-items.csv, evidence.json. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |