Claude Code for SDRs replaces the SaaS stack with a folder of markdown skills that run the workday end to end. The agent pulls a target list, scores it against your ICP, drafts personalized first touches, triages replies, and rolls a weekly report. You keep the calls and the message angle. The skills compound on disk.
The SDR workday is a stack of repeatable jobs
Most SDR tooling content treats AI as a copy assistant. That framing leaves 80 percent of the workday untouched. The honest read is that the job already breaks down into a small set of repeatable jobs: pull a list in the morning, qualify it, write the first touch, send and triage replies, keep the CRM clean, report on Friday. Five or six jobs. Same shape every week.
Claude Code is uniquely well placed here because it reads and writes real files, calls APIs directly, and lets you save the work as a skill you can edit, version, and rerun. That last property matters more than the model itself. A prompt you typed once is a memory. A markdown skill checked into a folder is an asset. The workday compounds when each job becomes a skill that gets a little sharper every iteration.
This piece walks the workday in order. Each section names a real skill, the MCPs it calls, and the output the SDR sees. The deep dive on the broader AI SDR landscape sits in the AI SDR tools field map for the layer above this. If you also run founder led sales or sit on the GTM ops side, pair this with Claude Code for sales for the closer view.
Morning list pull: the Crustdata skill
The first job of the morning is to produce today's target list. Most SDRs do this in a sales platform UI, export to CSV, and import to a sequencer. That is three tools and one tab switch for a job that takes the data API roughly three seconds.
The skill that owns this is crustdata-list-pull. The Crustdata MCP exposes people and company search directly into the Claude Code session, so the agent sends the search payload, reads the response, and writes the result as a CSV the next skill picks up. The operator types something like "pull 50 series A and B SaaS companies in the US that hired a first head of growth in the last 60 days, with the VP of marketing and the head of demand as targets." The skill turns the brief into a Crustdata query, runs it, dedupes against accounts you already booked, and writes today.csv to a date stamped folder.
The win is the audit trail. The CSV plus the query that produced it lives on disk. Tomorrow's run starts from the same file, the same filters, and the same exclusion list. The skill grows a notes.md next to it where the operator drops "this filter found too many agencies, tighten employee count to under 200." Next pull, the agent reads the note and applies the rule. The signal based list pull compounds because the rules live in markdown, not in a vendor UI you can't grep.
Qualification scoring: the qualify-leads skill
Sourcing produces volume. Qualification turns volume into the 15 accounts that deserve a real touch today. This is the job that templated AI SDRs do worst because the rule set is specific to your ICP and your customer evidence, not theirs.
The qualify-leads skill reads today.csv, hits an enrichment provider for missing fields, then scores each row against the ICP definition you keep in icp.md. The score is not a black box number. The skill writes a one line justification per row: "matches because Series B SaaS, hired first VP Sales last week, in your three closed industries." Or: "skip because the CEO is the only contact and the company is bootstrapped."
The point is that the rubric is editable. When the head of growth says "we are losing time on agencies, deprioritize anyone serving less than 50 employees," that lands as a one line edit in icp.md. The next run reads it. No vendor ticket. No config UI. The operator owns the playbook.
The output is today.qualified.csv with a column for the score and a column for the reason. The reps look at the top 15 rows, accept or override, and move on. Most teams who try this find they reject roughly a third of the model's top picks in week one, then closer to 10 percent by week three, because the model is reading their corrections back into the rubric.
Personalization at scale: the Unipile skill
This is where Claude Code earns its keep. A SaaS report cited in the MarketBetter SDR series puts AI assisted personalized emails at 2 to 3 times the reply rate of templates, produced about 5 times faster. The catch is that "personalization" only earns those numbers when it pulls from real context, not from a templated token.
The unipile-outreach skill, shipped through the Unipile outreach skill page, reads today.qualified.csv row by row. For each row it does three things. It pulls the prospect's last 30 days of LinkedIn activity through the Unipile API. It pulls the company's last hiring spike and recent press through Crustdata. It reads the rep's voice notes from voice.md so the tone is theirs, not generic.
Then it drafts. One LinkedIn note under 280 characters that opens on a real reference from the activity feed. One email under 90 words that hits a real signal and asks one question. Both written to a drafts/ folder, one file per prospect, each containing the inputs the agent used so the rep can audit the personalization in three seconds. The volume is whatever the day allows, typically 30 to 60 first touches if the rep is reviewing carefully.
Unipile is the right wiring here because it sits at the LinkedIn API layer rather than the browser layer. Pricing starts at around fifty five dollars a month for up to ten accounts as of June 2026, which is the right shape for a team that wants to hit several accounts from one orchestrator rather than per seat sequencer pricing that scales with the rep count.
Send and reply triage: the boundary that protects deliverability
Claude Code drafts. The sequencer sends. This rule is the one every honest SDR write up converges on, and the one every "fully autonomous AI SDR" pitch quietly breaks. There is a reason the SalesDuo critique of AI SDR replacements keeps landing: when the agent owns the send button, the operator loses the deliverability lever and the brand lever at the same time.
The skill that handles this is send-handoff. It picks up the drafts folder, runs a last pass deliverability and tone check, then writes the approved drafts to the sequencer. For email that means a clean push to Instantly or whatever sender owns the inbox. For LinkedIn that means a push to the Unipile send queue tied to the rep's account.
Reply triage is the other half. The reply-triage skill polls the unified inbox through the messaging APIs, classifies each reply into book / nurture / unsubscribe / objection, and writes a one line response suggestion per thread. Anything classified as "book" gets a calendar link and a draft confirmation. Anything classified as "objection" gets a draft rebuttal grounded in the call notes the rep keeps in markdown. The rep reads the queue, accepts or rewrites, hits send.
Two things make this practical instead of theoretical. The drafts stay reviewable because they sit as files, not as opaque API calls. And the rules live in the same voice.md the personalization skill reads, so the tone of the inbound response matches the tone of the outbound first touch.
CRM hygiene as a skill
Most SDRs spend a quiet 20 minutes a day cleaning the CRM and another 60 minutes a week reconciling what the sequencer told them with what their notes actually say. That work is a skill too.
The crm-hygiene skill talks to the CRM through the HubSpot MCP, reads the day's activity, and surfaces a short list every morning: duplicate contacts to merge, opportunities stuck in stage for over 21 days, accounts assigned to nobody, missing close dates on deals over a defined size. The rep gets a markdown list, fixes the top issues in the CRM UI in ten minutes, and moves on. The skill never writes destructive changes back to the CRM on its own. That guardrail keeps the system trustworthy.
The same MCP wiring lets the personalization skill read the prior conversation history before drafting a new touch. That single read prevents the most common cold email failure pattern: pitching someone you already talked to nine months ago.
Weekly report: the skill that reads the whole week
Friday afternoon is when most reporting gets faked. The numbers are real but the narrative is whatever fits the slide. A weekly report skill fixes that without adding work.
The weekly-report skill reads the week's drafts folder, the reply triage classifications, and the CRM activity log. It writes a short markdown report with the numbers (sent, replies, reply rate, booked, no shows), the patterns (which signal converted, which industry under indexed, which message angle worked), and the lessons (rules to add to icp.md or voice.md for next week). The report posts to a Slack channel. The rep reviews on Monday morning.
The win here is not the report. It is the feedback loop. The rules the rep changes on Monday land in markdown. The list pull skill picks them up Tuesday morning. The qualification skill applies them Tuesday afternoon. The personalization skill speaks in the new voice Tuesday night. The week compounds because the rules are markdown, not menu clicks.
The five skill starter kit for an SDR today
Here is the kit an SDR can install this week. Five skills, in order. Each one is a single markdown file that calls the right MCPs and writes its output to disk for the next skill to consume.
crustdata-list-pull: sources today's target list via the Crustdata MCP and writestoday.csv.qualify-leads: reads the list, scores againsticp.md, writestoday.qualified.csvwith reasons.unipile-outreach: drafts one LinkedIn note and one email per qualified row, writes todrafts/.reply-triage: classifies inbound replies through Unipile and the email API, drafts responses.weekly-report: rolls the week into a markdown report and posts it to Slack on Friday.
That is the floor. The crm-hygiene skill is the sixth and lands the day you wire the HubSpot MCP. Most operators get faster gains from adding the hygiene skill than from adding a sixth outreach channel.
Two cost notes worth doing the math on. A Claude Code Pro seat runs about twenty dollars a month as of June 2026. A typical full AI SDR replacement runs in the hundreds per month and ships a closed prompt you cannot edit. The cost gap is real but it is not the headline. The real difference is that the five files are yours. You can read them, change them, and walk them across tools when you change vendors.
What still needs a human, and what does not
Two parts of the workday do not belong to Claude Code, and pretending otherwise is what gave AI SDRs the deserved bad name. The first is the ICP definition and the message angle. That is first mile work. A model can synthesize past wins into a draft rubric, but the call about which ICP to bet the quarter on is the operator's. The second is the conversation: the discovery call, the negotiation, the customer success follow up that retains the deal. That is last mile work. The rep owns it.
Everything in the middle is fair game. Sourcing, enrichment, scoring, drafting, classifying, logging, reporting. This is what the AI sales agents emerging in 2026 actually do well when they are wired correctly. The dividing line is durable because it tracks where trust and judgment live. Trust lives in the relationship. Judgment lives in the strategy. Both stay with the human. The rest stays with the markdown files.
The other thing that still needs a human is the deliverability gut check. The agent can run a tone pass on every draft, but the rep is the one who notices that a particular sender warmed up too aggressively this week, or that LinkedIn is throttling invites on a Tuesday for no good reason. The orchestrator surfaces signal. The rep decides whether to send anyway.
Run the day from one Claude Code prompt
Pick one job and rebuild it as a skill this week. The fastest payback for most SDRs is the personalization step, because the time savings compound the moment the drafts folder lights up. The longest payback skill is the weekly report, because it pulls signal from every other skill, so it works best after the other four are running.
If you run a thin team or you sit on a one person GTM seat, the operator playbook for B2B lead generation gives you the framing for which playbook to layer on top of the kit. If you have a sequencer you like, keep it. Claude Code is not a sender. It is the orchestrator that drafts what your sender sends, scores what your CRM stores, and reads what your inbox returns.
The five files are the floor. The operator who keeps the rules in markdown and the conversations with humans wins the second half of 2026.
FAQ
Do I need coding skills to use Claude Code as an SDR?
No. The skills are plain English markdown files. You write the rules the way you would brief a junior rep, and the agent calls the APIs on your behalf. The only setup that touches anything like configuration is wiring the MCPs the first time, which is a short copy and paste from each provider's docs.
Can Claude actually send the cold email or does the sequencer have to?
The sequencer always sends. Claude Code drafts, runs a deliverability pass, and hands the approved copy to your sender of choice. Mixing the two breaks deliverability and brand voice in ways that are hard to recover. Every honest AI SDR write up converges on this rule.
Will AI generated cold emails hurt my deliverability?
Not on their own. Deliverability problems come from volume, list quality, and infrastructure, not from the source of the copy. If anything, draft level personalization protects you because each message is different enough to escape the worst of the spam fingerprinting filters. Warmup, dedicated domains, and a clean list still do the heavy lifting.
How is Claude Code different from a full AI SDR like 11x or Artisan?
A full AI SDR sells you a closed product where the prompt is the moat. You cannot read it, edit it, or move it. Claude Code with the five skill kit gives you the opposite shape: every prompt is a markdown file you can read, edit, version, and walk to another orchestrator. Same job, different ownership of the IP.
How does Claude Code compare to ChatGPT or Codex for SDR work?
ChatGPT is great for one off drafts inside a chat window. Codex is built for software engineering tasks. Claude Code sits in your terminal, reads and writes your files, calls APIs through MCP, and lets you save the work as a reusable skill. For repeatable SDR jobs the file based skill model is the difference between a memory and an asset.
What is an MCP and why does it matter for SDRs?
MCP is the Model Context Protocol. It is the wiring that lets Claude Code talk to a data provider or a CRM directly, instead of you copy and pasting CSVs. For SDRs it means the agent can read your HubSpot pipeline, run a Crustdata search, and check Unipile messages without you opening any of those tools. That is how the workday collapses into one prompt.