如何利用 Claude Code 构建一个你一个人就能运行的 GTM 团队

1Write team/prospector.py: run(memory, source, delivery, icp, sequences, weights, offline, dry_run).
2Sense new signals into the four buckets (job, social, company, funding) and write them to the shared
3memory. Cluster by account. For each account, judge it against the ICP and the full history with
4prompts/judge.md (Claude, temperature 0), falling back to a weighted-bucket heuristic that skips weak,
5off-ICP signals. For the ones that clear the bar, draft from the strongest trigger with prompts/draft.md,
6hand the draft to delivery (dry-run by default), and record the touch. Return ranked standup lines.

1Write team/researcher.py: run(memory, calendar, icp, offline).
2For each call on today's calendar, pull the account's full history from the shared memory and write a
3one-page brief with prompts/brief.md: what moved, what we already sent and how it landed, and the single
4best thing to open the call on. Ground every line in the history, never invent. Fall back to a plain
5template built from the latest signal when there is no API key. Return one standup line per call.

1Write team/sequencer.py: run(memory, delivery, sequences, offline, dry_run, min_age_days, max_age_days).
2Ask the shared memory for accounts whose last touch is a few days old and that have no reply or meeting
3on the books. Draft a light follow-up that adds one genuinely new angle, send it into the follow_up
4sequence, and record the touch. Leave fresh touches alone by honoring the min-age window. Return one
5standup line per account.

1Write team/recoverer.py: run(memory, calendar, delivery, sequences, offline, dry_run).
2For each recent no-show from the calendar, enrol it into the no_show_recovery sequence defined in
3config/sequences.yaml (four touches over a week, two channels, no blame in any step), record the touch,
4and return a standup line. The cadence lives in config, not in code.

1Write team/reporter.py: run(memory, weights, days, offline).
2Read the week's stats from the shared memory. Lead with the meeting count. Re-weight the four signal
3buckets by win rate, counting a meeting more than a reply, with a floor so no bucket ever drops to zero.
4Return the standup lines and the new weights for the next run. prompts/report.md writes the prose
5version when a key is set.

1ROLE  You are the prospector's judgment layer. For one account, decide whether a buying signal is worth
2acting on right now, and how.
3INPUT  { icp, new_signals: [{bucket, summary}], history, days_since_last_touch }
4SCORING  80-100 strong ICP fit and a high-intent signal (funding, or two clustered); 50-79 good fit,
5one solid signal; 20-49 weak fit or a lone low-intent signal; 0-19 off-ICP or noise
6RULES  Under 7 days since the last touch, prefer nurture or skip, never first_touch. If the signal is
7weak or off-ICP, score it low and skip; saying no is part of the job. why_now must quote the actual
8trigger, in words a rep could say to the buyer.
9OUTPUT (JSON only)  { "score": 0-100, "why_now": "...", "play": "first_touch|follow_up|nurture|skip", "rationale": "..." }

1ROLE  You write the opening message. The signal is the reason you are reaching out, and the message has
2to prove you noticed it.
3INPUT  { trigger, bucket, why_now, play, guardrails: {goal, must, never} }
4RULES  Open the first sentence on the trigger, never "Hi {{firstName}}". Connect the trigger to one
5problem you solve, in a single sentence. Close with one low-friction ask. Plain sentences, mixed length,
6no fake urgency, no em dashes, no buzzwords. If the message could have gone out unchanged last month, you
7skipped the trigger; start over.
8OUTPUT (JSON only)  { "subject": "6-9 words", "body": "3-5 sentences" }

1gtm-team/
2  team/     prospector · researcher · sequencer · recoverer · reporter
3  core/     memory · models · adapters · llm
4  prompts/  judge · draft · brief · report
5  config/   icp · sequences · signals · team
6  run.py    the morning standup