Deepline Gtm

GTM Meta Skill

Quick Start

npm install -g deepline
# Fallback for secure sandboxes: mkdir -p "$HOME/.local" && npm config set prefix "$HOME/.local" && export PATH="$HOME/.local/bin:$PATH" && npm install -g deepline --registry https://code.deepline.com/api/v2/npm/
deepline auth register --wait auto
deepline auth wait --timeout 120 # completes Cowork/browser approval; no-op if already connected
deepline auth status
deepline -h

CLI resolution

Run deepline commands bare — no pipes, redirection, 2>&1, command chaining, or backgrounding around them. The CLI already formats, truncates, and prints what you need; deepline billing usage | head reads as parsing and loses output.

Run deepline when it is available. If the shell reports that command is missing, use <workspace-root>/.deepline/runtime/bin/deepline (or the npm-created .cmd shim on Windows). If neither exists, follow https://code.deepline.com/INSTALL.md to set up Deepline.

Before the first Deepline fanout in a task, run deepline preflight --json as one standalone command and wait for it to finish. Never submit preflight beside another Deepline command. It combines health, authentication, and balance in one process and gives any automatic CLI update a serial boundary.

After preflight succeeds, prefix every Deepline command that may run concurrently with DEEPLINE_SKIP_SELF_UPDATE=1. This environment prefix is the only exception to the bare-command rule above. Serial commands may stay bare; the opt-out is required for every member of a parallel batch.

Ask for requirements, not implementation instructions. Requirements are the business outcome, target population, constraints, time horizon, requested destination, and any stated spend or authority boundary. The provider, tool, query shape, identifier recovery, filter expression, fallback order, and workflow structure are implementation decisions. Infer and execute the latter; do not turn them into a questionnaire. When a requirement is genuinely absent and materially changes the result or external action, ask one concise question with a recommendation. Otherwise state a reasonable assumption in the result and keep moving.

Decision-ready communication. The user should never have to infer the answer from a status update. When work produces people, companies, events, or rows that determine the next move, show those real records in a readable Markdown table first. Link a person's name to their verified LinkedIn profile when one was returned; do not hide the decision behind counts, summaries, or a generic Profile column. For choices, show the comparison; for copy, show the draft; for research, show the evidence that supports the conclusion.

Then make one plain-language recommendation based on what is visible. State the concrete boundary—who is in, out, or what changes—not a label such as “keep” or “refine.” Do not ask the user to design routine filters or choose plumbing. Keep validation, raw ids, feeds, and tool mechanics internal unless they change scope, cost, risk, confidence, or action.

When the user is calibrating, choosing a scope, prioritizing a list, or authorizing a change, use this complete envelope:

<the table, comparison, draft, or evidence>

Recommendation: <one concrete next state and why>.
Want me to use that, or adjust it?

End with Want me to use that, or adjust it? exactly. The user can say “yes” or name the adjustment. A calibration, scope, or prioritization response is incomplete without that final line, even when it starts no external action. Do not add a second question, a menu, or an implementation checklist. If no user decision is needed, state the outcome and stop.

Paid monitors. Before deployment show the recommended scope, live Deepline price, and delivery in the smallest useful shape. Check Slack first; recommend a real connected channel when one is available, otherwise offer Slack or the configured CRM. The monitor recipe covers consent, history, and similar-company scope.

Discovery order: companies first, then people. When the task requires finding contacts at companies matching criteria (portfolio, ICP, hiring signal), discover the company set first, then find people at each company. Do not start with broad people-search queries.

Named companies are enough to start. When the user gives company names but not domains, resolve each canonical company domain before asking them for anything. A domain is a recoverable identifier, not clarification debt. Carry the resolved domain and its official-page evidence into the downstream lookup; do not ask the user to paste a domain list merely because a later tool needs one. Read finding-companies-and-contacts.md for the identity gate and ambiguity handling.

Known companies + nuanced roles: qualify the real title roster first. For requests such as "AI leadership at Mount Sinai," "job titles at these companies," or "find the RevOps buyers at these accounts," read and follow recipes/find-qualified-titles.md: company_titles -> qualify exact roster titles -> deepline_native_search_contact with title_lists. Use Exa afterward for public-profile gaps and DropLeads last for supplemental database rows. Broad audience sizing remains a valid DropLeads use case.

CLI recovery

The SDK CLI is the supported CLI. If it is unavailable, run deepline update or reinstall it with npm install -g deepline@latest; do not switch CLI families.

2) Read the matching workflow before execution

SKILL.md routes; the matching doc supplies the execution contract. Read it before using a provider so its schema, sequencing, and known failure modes govern the run.

Routing rules — match your task to a doc and READ IT:

If you are hand-authoring enrich columns instead of using a native play, jump straight to the "Handmade step shape quick reference" section in enriching-and-researching.md. That section spells out the exact runtime contract for run_javascript, extract_js, result, and persisted matched_result.

Recipes: step-by-step playbooks for specific tasks (check before executing)

Read the matching recipe before executing. Follow its sequence; adapt it only when the request requires it.

Public/social source discovery, community-language pulls, pre-research source planning, or provider-coverage/cost comparison → use the standalone deepline-pre-research skill, not a recipe here. It owns X/Twitter, Reddit, Hacker News, Bluesky, and public-registry fanout plus the source-plan + Deepline-cost synthesis.

If none match, grep for more specific keywords: Grep pattern="<keyword>" path="<directory containing this SKILL.md>/recipes/" glob="*.md" output_mode="files_with_matches"

Data

• When the user hands you a CSV, run deepline csv show --csv <path> --summary first to understand its shape (row count, columns, sample values) before deciding how to process it.
• NEVER read a large CSV into context with the Read tool. Reading CSV rows into the conversation window exhausts context and produces zero output. This is the single most common failure mode.
• For row-by-row processing (enrichment, rewriting, research, scoring), use a Deepline play per §2.5: prebuilt if one fits, fork/wrap when close, author when not.
• To explore or understand CSV content without loading it, use deepline csv show --csv <path> --rows 0:2 for a two-row sample, or spawn an Explore subagent to answer questions about the data.
• Pilot before scale: slice the CSV (head -3 in.csv > pilot.csv), run the play on the slice, deepline runs export and inspect, then run the full file. Small-input exception: when an exact-fit prebuilt covers a small input (≤ ~25 rows) whose scope the user already stated, the full file IS the pilot — run it once.
• The pilot is never the deliverable. A task is not done until the FULL input has run and the result is exported to the exact requested output path: deepline runs export <run-id> --out "$FINAL_CSV". Finishing with only a pilot CSV, or an export under a play-derived name, is the single most common way to fail the task while feeling done.

Tools

For signal-driven discovery (investor, funding, hiring, headcount, industry, geo, tech stack, compliance), start with deepline tools search. Do not guess fields. Its syntax is deepline tools search [query] [--categories <categories>] [--search_terms <terms>] [--json]: provide a query, or at least one of --categories and --search_terms. The query is optional only for structured filtering. Use commas for multiple categories or search terms. There is no --prefix flag; include a provider name in the query when needed.

Search 2-4 synonyms, execute in parallel only after the standalone preflight:

DEEPLINE_SKIP_SELF_UPDATE=1 deepline tools search investor
DEEPLINE_SKIP_SELF_UPDATE=1 deepline tools search "crustdata investor"
DEEPLINE_SKIP_SELF_UPDATE=1 deepline tools search --categories company_search --search_terms "structured filters,icp"
DEEPLINE_SKIP_SELF_UPDATE=1 deepline tools search --categories people_search --search_terms "title filters,linkedin"

2.5) Plays are the surface

For row-by-row processing (per customer, per lead, per LinkedIn URL), run a Deepline play via deepline plays run. deepline enrich is deprecated — do not use or document it; when no play fits, author one.

1. Discover live, then run. deepline plays search <query> and deepline plays describe <name> — choose from the live catalog and its contract, never from memory. Search results include a runCommand and a cloneEditStarter for every prebuilt.

2. Prebuilt fits → run it. Batch prebuilts take a CSV directly: deepline plays run prebuilt/name-and-domain-to-email-waterfall-batch --input '{"csv":"leads.csv"}', then deepline runs export <run-id> --out "$FINAL_CSV".

3. Close but not exact → pull and edit it. Every prebuilt is forkable:

   deepline plays get prebuilt/<name> --source --out ./<name>.play.ts
   deepline plays check ./<name>.play.ts   # mandatory before running
   deepline plays run --file ./<name>.play.ts --input '{...}'
   

   If plays check fails on a missing local import, that prebuilt is multi-file — wrap it instead of forking: deepline plays bootstrap <family> --from <source> --using play:prebuilt/<name> --limit 5 --out workflow.play.ts.

4. No play fits → author one from scratch per recipes/deepline-plays.md: compose tools and other plays, map CSVs, add fallback logic and joins.

Results live in the Customer DB. Every batch play persists its dataset as a durable table: deepline db query --sql 'select * from "storage"."<table>" limit 20' --max-rows 20 --json (the run output names the table). Columns are the play's snake_case fields plus per-leg columns like email_result__hunter_email — the per-provider audit trail. Rerunning reuses filled cells instead of re-buying them; exports are projections of this table, so nothing is lost if a CSV goes missing.

The iterate loop — pilot, price, fix, then scale:

1. Run a few rows (slice the CSV or run 2-3 scalar inputs).
2. Read price and performance: deepline runs get <run-id> --full --json reports billing and per-step outcomes; the storage table's per-leg columns show which providers hit, missed, or erred.
3. Fix what the pilot exposed BEFORE scaling: a provider that misses or flakes on your segment gets dropped or reordered in a fork; wrong columns get a columns map; weak coverage gets a different route. Do not buy the same failure at full scale.
4. Run the full file, export to FINAL_CSV, and report to the user: rows delivered, coverage, observed credits, and what you changed after the pilot.

3) Core policy defaults

3.1 Definitions and defaults

GTM time windows, thresholds, and interpretation rules are defined in the Definitions section of enriching-and-researching.md.

Provider Playbooks

Provider-specific playbooks are bundled as separate reference files. Open the relevant playbook when provider-specific behavior, pricing, caveats, or payload conventions matter.

adyntel, affinity, ai_ark, akta, allegrow, amplemarket, apify, attention, attio, aviato, bettercontact, bigquery, bloomberry, bluesky, bounceban, browserbase, builtwith, clay, clickhouse, cloudflare, contactout, contextdev, crustdata, crustdata-v2, crustdata-v3, databricks, dataforseo, datagma, deepline_ip_to_company, deepline_native, deeplineagent, discolike, dropleads, emailbison, emailguard, enformion, enigma, exa, findymail, firecrawl, fireflies, forager, fullenrich, generic_http, gong, google_ads_audiences, hackernews, harvestapi, heyreach, hubspot, hunter, icypeas, instantly, intercom, ipqs, kernel, leadmagic, lemlist, limadata, linkedin_ads_audiences, lusha, meta_audiences, nooks, openmart, opensosdata, openwebninja, outreach, parallel, peopledatalabs, pipedrive, podscan, predictleads, prospeo, quickenrich, redshift, salesforce, salesforge, salesloft, scrapecreators, sentrion, serper, smartlead, snowflake, sumble, theirstack, trestle, twitterapi, upcell, versium, wiza, wizleads, zerobounce, zoho_crm, zoominfo

• Apply defaults when user input is absent.
• User-specified values always override defaults.
• In approval messages, list active defaults as assumptions.

3.2 Working directory — set up BEFORE any file writes

NEVER write files to /tmp/ or any absolute temp directory. Files in system /tmp/ are wiped on reboot — users permanently lose enriched CSVs, research outputs, and hours of paid enrichment work. This is a critical data-loss risk.

Set up a descriptive project-local working directory as your first action:

WORKDIR="deepline/data/<descriptive-task-slug>" && mkdir -p "$WORKDIR" && echo "$WORKDIR"

The slug must describe the task (e.g. deepline/data/yc-cmo-outbound, deepline/data/acme-email-waterfall). Do NOT use random names like mktemp generates — the user needs to find these files later. See enriching-and-researching.md for full details.

3.3 Output policy and User Interaction Pattern

• Always use a Deepline play for list enrichment or discovery at scale (>5 rows) — §2.5 routing. The run's play page lets the user inspect rows and rerun; send that URL.
• Even for company → ICP person flows, plays work: search and filter as part of the process, with providers like Apify to guide.
• Even when you don't have a CSV, create one and run the batch play against it.
• This process requires iteration; one-shotting via deepline tools execute is short sighted.
• In chat, send file/run links and render a decision table when rows inform a decision.
• Preserve lineage columns (especially _metadata) end-to-end. When rebuilding intermediate CSVs with shell tools, carry forward _metadata columns.
• Never overwrite a user-provided source CSV; write outputs to your working directory. Reruns of a play reuse completed cells by default.

See enriching-and-researching.md for deepline csv commands, pre-flight/post-run script templates, and inspection details.

3.4 Final file and run check (light)

• Keep one intended final CSV path: FINAL_CSV="${OUTPUT_DIR:-$WORKDIR}/<requested_filename>.csv"
• Before finishing: use the post-run inspection script pattern from enriching-and-researching.md. Run it once instead of separate checks.
• Checkpoint the deliverable. On multi-phase pipelines (companies → contacts → emails), write FINAL_CSV as soon as the first complete rows exist and overwrite it as later phases improve it. A timeout or crash must leave the best-so-far file at the requested path — intermediates under other names do not count as delivery.
• For a task that ran a Play, include its result. Give the exact FINAL_CSV path and play page link; do not invent one for monitor-only, research, or advisory work.

4) Credit and approval gate (paid actions)

This section's pilot, CSV preview, and full-run template governs enrichment, sourcing, and other row-processing runs. Monitor mutations use the workflow in recipes/deepline-monitors.md instead: inspect scope, reuse candidates, downstream actions/unknown consumers, and price before asking. Keep that approval decision-first: scope table, recommendation, live Deepline price, and one question. Because a monitor can incur variable future event charges, require explicit approval before every paid deploy, reactivate, or historical widening—even when the user stated the scope. A historical rung is empty only after its documented provider completion window; leave it intact while that window is pending. Ask only when a material requirement is missing or the check reveals an invalid or unaffordable configuration.

4.1 Required run order

1. Pilot on a narrow scope: a small CSV slice through the same batch play, or 2-3 scalar runs.
2. If the scope is NOT already approved (see below), request explicit approval.
3. Run full scope, report the pilot's cost and quality findings alongside the deliverable.

User-stated scope = already approved for bounded row-processing only. When the user's request itself states the full scope ("these 5 contacts", "~30 companies", "everyone in this CSV"), the request IS the approval: pilot to validate quality and provider choice, then complete the stated scope, export to FINAL_CSV, and deliver — reporting cost and per-provider performance with the result, not as a blocking question. This exception never applies to monitors; follow the monitor-specific consent gate above.

Stop and ask only when the scope is open-ended ("build me a big list"), the pilot reveals a problem worth a decision (low coverage, wrong matches, high cost per usable row), or projected spend exceeds a budget the user stated. Then present pilot results, projected cost, and the recommended route, and wait.

4.2 Execution sizing

• Use smaller sequential commands first.
• Keep limits low and windows bounded before scaling.
• For TAM sizing, a great hack is to keep limits at 1 and most providers will return # of total possible matches but you only get charged for 1.
• Prefer providers and plays that charge on returned results or successful hits when coverage is uncertain. If a provider bills per attempt/request/page, prove quality on a tiny pilot before letting it fan out.
• Stop after the pilot when the first rows show low usable coverage, wrong-person/company matches, missing getters, or high cost per usable row. Change route/provider order before buying the same failure at full scale.
• Do not depend on monthly caps as a hard risk control.
• Estimate play pricing before full scale: deepline plays list --show-cost, the play's describe output, and the pilot's observed cost from deepline runs get <run-id> --full --json. State the estimate in the approval message. deepline plays run has no cap flag, and the runtime-enforced --max-credits-per-run <credits> ceiling exists only on the deprecated legacy surface — never describe a play cap as enforced; the pilot plus stated estimate is the control.

4.2.1 Over-provision, then filter — never chase missing rows

When the user asks for N rows, start with ~1.4×N (e.g., 35 for 25). Every pipeline phase has natural falloff — contact search misses ~15-20% of companies, email waterfall misses ~5-10% of contacts. Fighting to complete the hard rows is almost always a waste: the companies that providers can't find contacts for are the same ones that won't have email coverage either.

Do this:

1. Pull more candidates than needed at the top of funnel.
2. Run the full pipeline (contacts → emails → outbound).
3. At the end, filter to the best N complete rows and deliver those.
4. Drop incomplete rows — don't retry or manually patch them.

Do NOT do this:

• Trim results to exactly N before running the pipeline.
• Spend turns retrying failed lookups with fallback providers, deeplineagent research passes, or manual patching.
• Run enrichment on all rows just to fill gaps in a few (especially broad deeplineagent research passes).

Provider coverage is a property of the company, not something you can overcome with more effort. Tiny startups with 5 people will have zero coverage across all providers — no amount of retrying changes that. Over-provision at the top and let incomplete rows fall off naturally.

4.3 Approval message content

Include all of:

1. Play or provider(s)
2. Pilot summary and observed behavior
3. Intent-level assumptions (3–5 one-line bullets)
4. CSV preview from the real pilot: the head of deepline runs export <pilot-run-id> --out
5. Credits estimate / range
6. Full-run scope size
7. Max spend cap (stated and monitored; no runtime-enforced play cap exists)
8. Approval question: Approve full run?

Strict format contract (blocking):

1. Use the exact four section headers: Assumptions, CSV Preview (ASCII), Credits + Scope + Cap, Approval Question.
2. If any required section is missing, remain in AWAIT_APPROVAL and do not run paid/cost-unknown actions.
3. Only transition to FULL_RUN after an explicit user confirmation to the approval question.
4. run_javascript is the non-AI path. ai_inference is for general classification/structured reasoning, and deeplineagent is for context gathering / web research / signal extraction.

Approval template:

Assumptions

- <intent assumption 1>
- <intent assumption 2>

CSV Preview (ASCII)
<paste verbatim pilot output: the runs export head>
Credits + Scope + Cap

- Provider: <name>
- Estimated credits: <value or range>
- Full-run scope: <rows/items>
- Spend cap: <cap>
- Pilot summary: <one short paragraph>

Approval Question
Approve full run?

4.4 Mandatory checkpoint

• Must run a real pilot against the exact CSV intended for the full run: a small slice through the same batch play.
• Must include the pilot output preview verbatim in approval.
• If pilot fails, fix and re-run until successful before asking for approval.
• Ask for approval in chat after the pilot. Include the row count, estimated credits, and a small ASCII preview so the user can approve or redirect without opening another surface.

4.5 Billing commands

deepline billing balance  # Show current credit balance
deepline billing usage    # Show recent billing activity and grouped recent usage
deepline billing limit    # Show the current monthly billing cap

When credits are zero or unavailable, stop paid work and ask whether the user wants to add Deepline credits. If the balance or failure output includes a recovery object, quote its top_up_command and checkout_command exactly, including --json and --no-open; do not run them until the user approves. Do not hardcode a USD-to-credit exchange rate in the skill. Use live billing, pricing, or tool output when quoting credit costs.

5) Provider routing (high level)

Quick-reference summary only — the Section 2 sub-doc you already read is the authority.

• Search / discovery → You MUST have finding-companies-and-contacts.md open. It contains the parallel execution patterns, provider filter schemas, and provider mix tables. Start with deepline tools search <intent> and execute field-matched provider calls in parallel; when the deepline-list-builder subagent is available, use subagent-based parallel search orchestration as the preferred pattern. Use deeplineagent only for synthesis or ambiguity resolution after the direct discovery path is exhausted.
• Enrich / waterfall / coalesce → You MUST have enriching-and-researching.md open. It routes each scenario to a play and shows the deepline plays run invocation, plus waterfall patterns and coalescing logic. Do not restate play internals from memory; treat the play itself as the source of truth for exact provider order and gating.
• Custom signals / messaging → Read enriching-and-researching.md (custom signals section). Use run_javascript for deterministic transforms/template logic and deeplineagent for AI work. Start from prompts.json.
• Verification → leadmagic_email_validation first, then enrich corroboration.
• LinkedIn profiles, company employees, posts, comments, and reactions -> Prefer Deepline's native HarvestAPI provider. Use the documented harvestapi_* names as starting hints, then run deepline tools describe <operation> --schema-only before execution; broad tool search can be noisy. Use Apify only when HarvestAPI does not expose the required LinkedIn surface.
• For phone recovery, read enriching-and-researching.md and follow the notes/provider guidance there rather than relying on deleted numbered sections.

Before hand-rolling any pipeline a prebuilt might cover, deepline plays describe the candidate play and either use/wrap it or state the contract mismatch in one line. Silently bypassing a fitting prebuilt is a routing failure.

Provider path heuristics:

• Broad first pass: direct tool calls for high-volume discovery.
• Quality pass: AI-column orchestration with explicit retrieval instructions.
• For job-change recovery: prefer quality-first (crustdata_person_enrichment, peopledatalabs_*) before leadmagic_* fallbacks.
• Never treat one provider response as single-source truth for high-value outreach.

6) Additional notes

Critical: keep writing-outreach.md workflow context active when running any sequence task. It is not optional for ICP-driven messaging.

Operational troubleshooting: rate limits and CLI health

• Use Deepline plays for heavy row-by-row work whenever possible. The runtime has built-in rate-limit handling (adaptive retries/backoff) for standard upstream limits. If you are building a homegrown script, assume it does not include the same automatic protection unless you explicitly implement it.
• If enrichment or CLI behavior is unstable, update the CLI and reinstall the Deepline skills:

deepline update
deepline skills

Sites requiring auth: Don't use Apify. Tell the user to use Claude in Chrome or guide them through Inspect Element to get a curl command with headers (user is non-technical).

1. If user provides actor ID/name/URL: use it directly.
2. If not, search deepline tools describe apify_run_actor_sync for the actor id, or try deepline tools search.
3. If not present, run discovery search.
4. Avoid rental-priced actors.
5. For LinkedIn posts and engagers, use the native HarvestAPI operations first: inspect harvestapi_search_posts, harvestapi_get_post, harvestapi_get_post_reactions, and harvestapi_get_post_comments. Use supreme_coder/linkedin-post only when the native provider does not cover the requested shape. Avoid silentflow/linkedin-posts-scraper-ppr and alizarin_refrigerator-owner/linkedin-post-scraper unless the user explicitly asks for them.
6. Pick high rating plus high usage/run count; when tied, choose best evidence-quality/price balance.
7. Honor operatorNotes over public ratings when conflicting.

deepline tools execute apify_list_store_actors --input '{"search":"similarweb traffic scraper","sortBy":"relevance","limit":20}'
deepline tools execute apify_get_actor_input_schema --input '{"actorId":"radeance/similarweb-scraper"}'

7) Feedback & session sharing

7.1 Proactive issue reporting (mandatory)

Do not wait for the user to ask. If there is a meaningful failure, send feedback proactively using deepline feedback send.

Trigger when any of these happen:

• A provider/tool call fails repeatedly.
• Output is clearly wrong for the requested task.
• A CLI/runtime bug blocks completion.
• You needed a significant workaround to finish.

Run once per issue cluster (avoid spam), and include:

• workflow goal
• tool/provider/model used
• failure point and exact error details
• reproduction steps attempted

deepline feedback send "Goal: <goal>. Tool/provider/model: <details>. Failure: <what broke>. Error: <exact message>. Repro attempted: <steps>."

7.2 End-of-session consent gate (mandatory)

After the substantive result is resolved, ask exactly one Yes/No question in a separate message—never beside a table, recommendation, approval request, or unresolved user decision. Do not ask it after a terminal no-change outcome: Change: none — I’m leaving it as is. must remain the final line of that response

…(truncated)