Growth Acquisition for DeepSeek Harness

中文 · English

dsh-growth is a local-first DeepSeek Harness bundle for evidence-backed user growth and customer acquisition analysis.

It covers AARRR funnels, activation, retention cohorts, referral loops, MRR bridges, CAC/LTV/payback, HADI experiments, RICE prioritization and WBR/MBR reports for Markdown, CSV and JSONL data.

User and company pain points

Growth work often breaks down in the gap between data, decisions and execution:

Pain point	Required capability
Growth data is scattered across notes, event exports, revenue sheets and team documents.	Read local Markdown, CSV, JSON and JSONL with a consistent analysis flow.
Teams use different definitions for activation, retention, CAC, LTV and MRR.	Make metric definitions, fields, periods, sources and caveats explicit.
Funnel dashboards show where users drop, but not what to investigate next.	Identify bottlenecks, segment differences, evidence gaps and next checks.
Ideas become long backlogs without a falsifiable hypothesis or owner.	Turn opportunities into HADI experiments with guardrails and RICE/ICE scores.
Weekly and monthly reviews are repetitive, disconnected from experiments and hard to audit.	Generate evidence-linked WBR, MBR, QBR and experiment-review Markdown.
Sensitive customer data should stay inside the company's knowledge boundary.	Keep analysis local-first with path limits, warnings and guarded writes.

Application scenarios

Scenario	How dsh-growth is used
New product or PMF discovery	Audit JTBD, ICP, PMF Survey, North Star and evidence readiness.
Customer acquisition	Compare acquisition, activation and revenue conversion by channel and segment.
Onboarding optimization	Locate the activation bottleneck and create a measurable HADI experiment.
Retention improvement	Build day/week/month cohorts, inspect lifecycle states and compare user segments.
SaaS or subscription monetization	Reconcile MRR movements and calculate ARR, NRR, CAC, LTV and payback.
Growth operating cadence	Produce weekly/monthly reviews with findings, decisions, caveats and next actions.

Included tools

Tool	Purpose
growth_onboarding	Run a read-only readiness check across strategy notes and datasets; show ready, partial, missing and unsupported methods
growth_doctor	Check the local workspace and summarize dataset health before analysis
growth_profile_dataset	Infer fields, coverage, date range and data-quality warnings without raw rows
growth_review	Start from a business goal and orchestrate profiling, analysis, bottleneck and next actions; paths may be omitted for local auto-discovery
growth_audit_note	Audit one growth note for JTBD, PMF, North Star, AARRR and evidence quality
growth_audit_vault	Scan a local knowledge base for growth-document gaps
growth_funnel_analyze	Analyze AARRR-style event funnels by channel and segment
growth_cohort_analyze	Analyze retention cohorts and lifecycle states
growth_economics	Calculate MRR bridge, CAC, LTV, NRR and payback
growth_diagnose	Diagnose a growth change and rank evidence-backed hypotheses
growth_experiment	Create a HADI experiment card and RICE/ICE score
growth_prioritize	Rank growth opportunities with RICE or ICE
growth_report	Generate WBR, MBR, QBR or experiment-review Markdown
growth_apply	Preview or guarded-write Markdown under the configured root

Quick start

Zero-threshold path

You do not need to know the tool names, AARRR definitions, field mappings or which dataset to open first. You only need:

1. A business question, such as “why did activation fall?” or “which acquisition channel should we scale?”
2. A configured local growth root, or a file path if you already know the relevant file.
3. Permission to read local data; the plugin does not upload your vault.

For a new project, start with a readiness check:

Run a growth onboarding check for my configured root.
Tell me what is ready, partial, missing and not supported, and give me only the top two gaps to fix next. Do not write files.

It checks growth notes and local CSV/JSON/JSONL datasets without returning raw user rows. It also shows which classic methods are detected in the project, which are only available as audits or templates, and which require an external system. If you already know the goal and want analysis immediately, use growth_review instead.

The shortest first review request is:

Run a growth review for the goal "improve activation" using the best available data under my configured root.
Show me which files you selected, what is missing, the biggest bottleneck and the next check. Do not write files.

If no paths are supplied, growth_review scans the configured root, profiles supported CSV/JSON/JSONL files, selects the most analysis-ready event and MRR sources, and explains the selection. If there are several candidates, confirm the selected files before making a budget or product decision.

Install the plugin into a DeepSeek Harness profile:

npx --yes @deepseek-ai/dsh plugin --profile growth add dsh-growth
npx --yes @deepseek-ai/dsh --profile growth --dump-config

If the Harness host is not running yet, start its Web UI first:

npx --yes @deepseek-ai/dsh web

Then open http://127.0.0.1:3080 in your browser and use the growth profile. DeepSeek Harness's official repository documents this Web UI entry point; the host is in developer preview, so its setup commands may evolve.

If dsh is already on your PATH, the equivalent short form is dsh plugin --profile growth add dsh-growth.

Configure the plugin through the host. A minimal configuration is:

defaultRoot: ""
reportDir: ".dsh-growth/reports"
defaultCurrency: "CNY"
defaultTimezone: "Asia/Shanghai"

Then use the tools from the conversation. Typical requests are:

Run a growth review for the goal "improve activation" using the best available data under my configured root; show which files you selected and what is missing.
Run a growth review for the goal "improve activation" using events.csv; tell me what is missing before making a recommendation.
Audit growth-plan.md for PMF, North Star, AARRR metrics and evidence gaps.
Analyze events.csv as an AARRR funnel and compare channel and segment performance.
Analyze mrr.csv for MRR Bridge, NRR, CAC, LTV and Payback using a gross margin of 0.8.
Turn the largest activation bottleneck into a HADI experiment and score it with RICE.
Generate this week's WBR as Markdown; do not write a file yet.

The recommended conversation flow

Use these six requests in order when you are new to the plugin:

1. Run a growth onboarding check; tell me what is ready, partial, missing and not supported.
2. Review the goal "improve activation" and tell me what is missing before giving a conclusion.
3. Break down the bottleneck by channel and segment; separate evidence from hypotheses.
4. Turn the highest-leverage hypothesis into a HADI experiment with a primary metric and guardrails.
5. Score the experiment with RICE and show which inputs are estimates rather than observed facts.
6. Generate this week's WBR; preview only and do not write a file.

You can replace the goal and the file names without changing the workflow. Advanced users may call the individual tools directly, but that is optional.

Tool results use a stable envelope with ok, data, warnings, assumptions, lineage and nextActions. Read warnings and lineage before using a number in a decision.

Input conventions

Event data should use user_id, event and timestamp, with optional channel, segment, plan, revenue and currency fields. MRR data should use period, type, amount, customer_id, active_customers and spend. Supported movement types are new, expansion, reactivation, contraction, churn and churned. The goal-oriented review recognizes common English and Chinese event values such as signup / 注册, activated / 激活, active / 活跃, invited / 邀请 and paid / 付费.

For the first review, eventPath and economicsPath can be omitted. growth_review scans the configured local root, selects the most analysis-ready event and MRR files, and records the selected sources in assumptions, warnings and lineage. If more than one file is suitable, confirm the selection before using the result for a budget or product decision.

Minimal data examples

An event file can start with only these three fields:

user_id,event,timestamp
u001,signup,2026-08-01T09:00:00Z
u001,activated,2026-08-01T09:20:00Z
u001,active,2026-08-08T09:20:00Z

For MRR and acquisition cost analysis, add the fields you actually have:

period,type,amount,customer_id,active_customers,spend,currency
2026-08,new,1000,c001,20,5000,CNY
2026-08,expansion,200,c002,20,,CNY
2026-08,churned,100,c003,20,,CNY

Do not fabricate missing columns. The plugin will keep dependent metrics unavailable and explain what needs to be added.

How to read a result

Every tool returns the same outer structure:

Field	Meaning	What you should do
ok	Whether the tool completed	Stop if false; correct the reported error
data	Analysis or Markdown report	Read only after checking the other fields
warnings	Data-quality risks and limitations	Treat as part of the result, not as decoration
assumptions	Defaults or automatic source selection	Confirm before using the result for a decision
lineage	Source files, fields and time windows	Use it to trace every important number
nextActions	Concrete next checks or actions	Pick one owner and one decision date

null means “not available or not trustworthy with the supplied data”; it is not zero. For example, missing spend leaves CAC and payback unavailable, and missing beginning MRR leaves first-period growth and NRR partial.

Safe write workflow

Reports are returned as Markdown and are not written automatically. When updating an existing Markdown file:

1. Call growth_apply with the complete Markdown content and confirm=false to preview.
2. Review the preview and call it again with the same content and confirm=true only after approval.

Writes stay inside defaultRoot and use a version guard to avoid overwriting concurrent edits. Read warnings before interpreting analytical results; missing amounts, spend, active-customer counts and beginning MRR remain unavailable instead of being silently treated as zero.

Troubleshooting without technical knowledge

What you see	What it means	What to ask next
No usable dataset found	The configured root has no supported or recognizable data	Check my configured growth root and tell me what file and field is missing.
Multiple event/MRR files selected	The plugin found more than one plausible source	Use events-prod.csv for events and mrr-2026.csv for economics.
Fewer than two funnel stages	Event names were not recognized or the event file is incomplete	Profile events.csv and show the event values; use these explicit stages: signup=注册, activation=激活.
CAC/LTV/Payback is null	Spend, active customers, gross margin or churn evidence is missing	List the exact inputs needed to calculate CAC, LTV and Payback.
NRR or MRR growth is partial	Beginning MRR or movement amounts are missing	Tell me which periods and movement rows prevent a complete MRR bridge.
Write was rejected	The target is outside the root, not Markdown, or changed since preview	Preview the report again and show the safe path under my growth root.

If the installed plugin does not expose growth_review, start a new Harness/Codex thread after reinstalling the plugin so the new tool manifest is loaded.

Defaults

The plugin is configured for a local knowledge base, a report directory, a default currency and a timezone. These values are supplied by the host configuration and can be adjusted for each environment.

The plugin is local-first. It does not upload a vault or call an external API unless an optional connector is explicitly added in a later phase.

Development

pnpm install
pnpm run typecheck
pnpm run lint
pnpm test
pnpm run build

The plugin follows the standard Cordis bundle contract: it exports apply(ctx), injects tools and fs, and registers model-facing tools through the normal tool pipeline.

Methodology

• Jobs to Be Done for problem and customer context.
• Sean Ellis PMF survey as a heuristic gate.
• North Star Metric and driver tree.
• AARRR funnel plus growth loops.
• Cohort retention and lifecycle analysis.
• HADI experiment cards.
• RICE/ICE opportunity prioritization.
• MRR bridge and unit economics.

License

MIT. See LICENSE.