robomotionio Sequence Performance
Deterministic pull + rate math; you, the agent, classify every reply, grade the copy, assess lead quality, and write the recommendations. Metrics stay computed (trustworthy).
v2.0.0 — depth upgrade.
analyze_sequence.pynow computes per-touch & per-variant open/reply/positive rates, compares against a built-in per-segment reference table (SMB / MidMarket / Enterprise × cold/warm), normalizes pre-tagged replies into a canonical reply-category taxonomy, and extracts copy-grading signals (subject length, CTA, personalization tokens, spam triggers). This SKILL.md carries the full benchmark tables, the reply taxonomy, the copy rubric, and the output skeleton.
When to use
- "How's my campaign doing", "sequence performance", "campaign review", "email analytics".
- "Why isn't my campaign working", "review my email results" (after 7+ days of data).
How to run
Step 1 — pull the campaign bundle (deterministic)
python3 ${SKILL_DIR}/scripts/fetch_campaign.py --tool instantly --campaign "Q1 Cold" \
--output ${WORKSPACE}/bundle.json
Tools: instantly, lemlist, or keyless csv:
python3 ${SKILL_DIR}/scripts/fetch_campaign.py --tool csv --campaign "Q1 Cold" \
--metrics-csv ${WORKSPACE}/metrics.csv \
--replies-csv ${WORKSPACE}/replies.csv \
--copy-csv ${WORKSPACE}/copy.csv \
--output ${WORKSPACE}/bundle.json
CSV shapes: metrics = touch,variant,sends,opens,replies,bounces; replies =
date,from,text; copy = touch,variant,subject,body. The API tools return overall
totals; for per-touch/per-variant depth use the CSV path or supplement.
Step 2 — compute rates + confidence (deterministic)
python3 ${SKILL_DIR}/scripts/analyze_sequence.py --bundle ${WORKSPACE}/bundle.json \
--outreach-type cold --segment SMB \
--benchmark-reply-rate 0.03 --benchmark-open-rate 0.40 \
--output ${WORKSPACE}/metrics.json
Computes overall + per-touch + per-variant open/reply/positive/bounce rates, marginal
reply share per touch, vs-passed-benchmark deltas, a per-segment comparison (pick
--segment SMB|MidMarket|Enterprise), a normalized reply taxonomy (from any pre-tagged
reply categories in the bundle), copy signals per touch/variant, and a variant
confidence label (insufficient <50 / low / moderate / significant 250+). data_quality
flags missing open tracking / variant data / reply text / positive tracking / copy.
Built-in segment benchmarks (cold; the bar the script compares against)
| Segment | Open | Reply | Positive-reply | Bounce ceiling |
|---|---|---|---|---|
| SMB | 45% | 5.0% | 1.5% | 3% |
| MidMarket | 40% | 3.5% | 1.2% | 3% |
| Enterprise | 35% | 2.5% | 0.8% | 3% |
(Warm/nurture rows run higher — opted-in lists clear ~2–3× the reply bar.) segment_comparison
emits a verdict per metric: above (≥1.2× bench) · at (0.8–1.2×) · below (<0.8×).
Enterprise opens lower but each reply is worth more — weigh positive_rate, not raw reply.
Reply-category taxonomy (the script normalizes; you refine from text)
Canonical buckets the script maps any pre-tagged reply.category into — classify every
raw reply you read into the same buckets so the counts and your narrative agree:
| Bucket | What lands here | What it means |
|---|---|---|
positive |
interested / meeting / demo / "yes" | The win metric. Track positive_rate, not reply_rate. |
objection |
not relevant · budget · competitor · already have | >20% "not relevant" = targeting problem, not copy. |
not_now |
timing / next quarter / circle back | Nurture, don't kill — sequence them later. |
referral |
wrong/right person, forwarded | A win in disguise — re-route to the named person. |
out_of_office |
auto-replies | Noise; exclude from reply-quality math. |
unsubscribe |
remove / opt-out / GDPR | List/targeting hygiene + compliance flag. |
other |
anything unmatched | Read and re-bucket manually. |
Copy-grading rubric (signals → letter grade)
The script extracts signals; you assign the A–F grade. Targets for cold outbound:
| Signal | Healthy target | Read |
|---|---|---|
subject_word_len |
3–7 words | Long/clickbait subjects tank opens. |
body_word_count |
50–125 words | Walls of text kill replies. |
has_cta |
true | One clear, low-friction ask (a question, a 15-min). |
personalization_token_count |
≥1 real token | Zero tokens on cold = spray-and-pray. |
spam_trigger_count |
0 | Each trigger word risks the spam folder (deliverability). |
question_count |
1–2 | A question invites a reply; none = a monologue. |
Grade each subject + body A–F on hook, value prop, proof, personalization, CTA, filler, and sequence progression (does touch 2 add a new angle or just "bumping this"?).
Step 3 — analyze + write the report (you, the agent)
- Reply analysis — read
bundle.jsonreplies; classify each (positive / meeting / warm / objection-timing/budget/competitor/relevance/authority / not-interested / OOO / referral / question); extract objection + positive-signal patterns; preserve exact objection language. Assign a reply-quality score. - Copy quality — grade subjects + bodies A-F (hook, value prop, proof, personalization, CTA, filler, sequence progression).
- Lead quality — compare titles/industry/seniority/size to ICP; infer targeting issues from reply patterns (e.g. >20% "not relevant" = targeting, not copy).
- Diagnose by pattern: low open = subject; good open + low reply = body; high bounce = list hygiene; low deliverability = SPF/DKIM/DMARC. Don't kill a variant under 100 sends. Write the exec summary (grades + top 3 actions), the detailed sections, and a kill list.
Diagnostic decision table:
| Symptom (from metrics) | Root cause | Fix |
|---|---|---|
open_rate below segment, reply ok |
Subject line / sender reputation | Rewrite subjects (3–7 words); warm up the domain. |
open_rate at/above, reply_rate below |
Body / offer / CTA | Tighten body to 50–125 words; one clear ask; stronger proof. |
positive_rate low but reply_rate ok |
Wrong audience or weak offer | Re-check ICP fit; reply taxonomy heavy on objection/not relevant = targeting. |
bounce_rate over ceiling |
List hygiene | Re-verify the list (email-validator); pause sends. |
Reply taxonomy heavy unsubscribe |
Targeting/compliance | Tighten list; review consent + frequency. |
Variant gap with confidence < significant |
Not enough data | Don't kill it — keep running to ≥250 sends. |
A later touch has high marginal_reply_share |
Follow-ups are doing the work | Keep the sequence long enough; don't stop at touch 1. |
Output skeleton:
# Sequence Performance — [Campaign]
## Executive Summary
- Open X% / Reply X% / Positive X% / Bounce X% — vs [segment] benchmark: [verdicts]
- Copy grade: [A–F overall] · Confidence: [label]
- **Top 3 actions:** 1) … 2) … 3) …
## Metrics vs Benchmark (overall + per-touch + per-variant table w/ confidence)
## Reply Deep-Dive (taxonomy counts + % + verbatim objection language)
## Copy Assessment (per touch/variant: signals → A–F grade + the rewrite)
## Lead-Quality Assessment (ICP fit + targeting verdict from reply patterns)
## Recommendations & Kill List (what to fix, what to pause, what to scale)
Outputs
sequence-performance-[YYYY-MM-DD].md— exec summary, detailed metrics (vs benchmark, per-touch, per-variant w/ confidence), reply deep-dive, copy assessment, lead-quality assessment, prioritized recommendations + kill list. Workspace + Agent Teams attachment.
Credentials / env
- Required: none for the CSV path (both scripts are keyless). No LLM key — the classification, grading, and recommendations are your job as the agent.
- Optional: if an outreach-tool key is set (
INSTANTLY_API_KEYorLEMLIST_API_KEY) →fetch_campaign.pypulls the campaign via API; if not → the keyless--tool csvpath (default).
Notes & edge cases
- Degrades gracefully: minimum viable = sends + reply count + copy text. Missing reply
text drops classification; missing variant data drops A/B analysis; missing open
tracking drops open-rate analysis — the rest still runs (see
data_quality). - Reply analysis is where the insight is — metrics say WHAT, replies say WHY.
- Pure reasoning over the user's tool data — no scraping, no paid API.