How to Build a Cold Email Campaign with Codex
Use Codex to construct the research, enrichment, validation, scoring, personalization, review, and measurement layers of a cold-email workflow.
Codex can make a cold-email workflow faster to build, but speed is not the same as permission or quality. The strongest use of an agent is to construct and test the system around a campaign: data contracts, enrichment logic, validation, scoring, review queues, integrations, and measurement.
Keep the human accountable for audience policy, legal basis, suppression rules, claims, tone, and the final decision to launch. Use the agent to reduce repetitive implementation work and make the workflow easier to inspect.
What Codex should build
| Layer | Agent contribution | Human decision |
|---|---|---|
| ICP | Translate criteria into explicit filters and tests | Approve who should and should not be targeted |
| Data | Source, map, enrich, deduplicate, and preserve provenance | Approve sources and acceptance rules |
| Contact | Find and validate business contact paths | Set risk and suppression policy |
| Personalization | Draft from approved evidence and templates | Approve claims, tone, and examples |
| Execution | Prepare imports and integration payloads | Authorize launch and sending limits |
| Measurement | Join delivery and outcome data back to the workflow | Interpret results and change strategy |
Step 1: write the campaign contract
Give Codex a concrete objective, audience definition, exclusions, required evidence, contactability rule, review policy, destination schema, and success metric. Add explicit prohibitions: no fabricated facts, no sensitive-trait inference, no automatic launch, and no records without provenance.
Goal: book qualified discovery calls with US B2B security companies
ICP: 50–500 employees, defined categories, approved regions
Roles: VP Security, VP IT, Head of Infrastructure
Required: source URL, company evidence, verified work email
Exclude: customers, open opportunities, unsubscribes, competitors
Launch: human approval required
Success: positive replies and qualified meetings, not send volumeStep 2: create the source table
Start from gtmcli's people or company data, a reviewed CSV, a webhook, an API, or a suitable Apify Actor. Keep the source URL, source type, collected time, and raw identifier attached to the row.
Step 3: enrich in a deliberate order
Enrich the company before the person when company context determines fit. Then identify target roles, discover contact information, and validate it. Put cheap, decisive checks before expensive research steps.
- Normalize company domain and deduplicate against customers and prior campaigns.
- Add firmographic and source-specific evidence.
- Identify people who match the approved role rule.
- Find a work email through a documented waterfall.
- Validate the email and isolate catch-all or unknown results.
Step 4: make scoring explainable
Ask Codex to implement a score from visible inputs, not to emit a mysterious number. Store the reason alongside the score. If a criterion cannot be explained or tested, remove it from automated qualification.
Step 5: personalize from evidence
Personalization should use approved, source-backed facts that are relevant to the business problem. The agent can transform evidence into a draft, but it should not invent observations, pretend to have personal familiarity, or use sensitive personal information.
Write one concise opening sentence using only evidence_source
and evidence_summary. If the evidence is weak or unrelated to
the campaign problem, return NEEDS_RESEARCH. Do not infer intent.Step 6: validate the finished row
Before export, require a stable company key, approved role, source evidence, valid contact status, score reason, suppression check, and message-review status. Send incomplete records to a review queue rather than filling gaps with guesses.
Step 7: prepare—not launch—the campaign
Codex can map accepted records into an Instantly, Smartlead, HeyReach, EmailBison, or another supported destination format. Keep the actual launch behind human approval, especially during a new campaign or when the source and scoring logic have changed.
Step 8: close the loop
Join bounces, replies, positive replies, qualified meetings, suppressions, and complaints back to the original record. Evaluate by segment and source. A campaign with fewer sends and better accepted-record quality may be the stronger system.
A prompt for Codex
Build this as an editable gtmcli workflow. Start from the approved account source, enrich company and target roles, find and validate work emails, calculate an explainable ICP score, draft only from cited evidence, and queue accepted rows for human review. Do not launch messages or write to the CRM without approval.
Next steps
Connect the agent using the gtmcli MCP guide.
Design reliable triggers and review states with the GTM automation guide.
Clean an imported audience with bulk email verification.