claude-for-revenue-intelligence

claude-for-revenue-intelligence is a small, readable agent harness for schema-driven revenue intelligence. It is designed to be forked and specialized through skills: the base stays motion-agnostic, while each skill binds schema slots, agent roster, persona plugin defaults, cookbook set, connector bindings, and motion-specific theory constants.

Nothing here is a vendor product or a commercial benchmark. It is a reference implementation meant to be read end-to-end and adapted.

The Pattern

The harness follows this flow:

schema -> agents -> persona plugins -> cookbooks -> connectors

• Schema: column contracts supplied by the active skill.
• Agents: small modules that populate or watch schema slots.
• Persona plugins: role-shaped views over the same skill-bound schema.
• Cookbooks: end-to-end workflows that compose agents and plugins.
• Connectors: read/write adapters for systems of record, defaulting to read-only.

The loader at skills/loader.py reads CLAUDE.local.md, selects the active skill, and exposes the bindings agents need. If no local profile exists, the loader falls back to enterprise-account-based.

How To Specialize

Run the cold-start interview in QUICKSTART.md and choose an installed skill, or fork one of the examples under examples/forks/.

A specialization lives in a folder with a SKILL.md and its own schema contracts. The base harness should not hardcode motion assumptions; it should read schema slots and theory constants from the active skill.

Executable helpers make that loop concrete:

python tools/inspect_skill.py
python tools/cold_start.py --non-interactive --skill enterprise-account-based --profile-path CLAUDE.local.md
python tools/new_skill.py plg-self-serve skills/my-plg-motion

inspect_skill.py prints the active bindings. cold_start.py writes a local practice profile. new_skill.py copies an example fork into a new skill folder.

Reference Skill

The default reference skill is:

• skills/enterprise-account-based/

This is the original six-part model, relocated without changing its semantics. It remains the default behavior for current agents, plugins, demos, and evals.

Example Forks

Stub overlays demonstrate that the harness can bind very different motions:

• examples/forks/finserv-enterprise/: current six-part model plus regulatory filings emphasis.
• examples/forks/plg-self-serve/: product usage, activation events, and expansion signals.
• examples/forks/healthcare-patient-acquisition/: referral authority, episode telemetry, and outcome evidence.

The examples intentionally do not implement full agents.

Reference Skill Slots

The enterprise-account-based skill supplies these schema slots:

• signature_authority: actual signatory data from public SEC filings and contract corpora; pen-on-paper authority, not titles.
• persona_graph: relationship and influence map per account; who decides, who blocks, who champions, and how those edges are observed.
• funnel_telemetry: first-contact date, touch count, days to close, opportunity counter, outlier filter, and the anti-qualification ratio. The ratio formula and thresholds are skill-level theory constants.
• outcome_telemetry: post-implementation news, contract diffs, renewal signals; the slot that tells you whether the deal you closed actually became a customer.
• conversation_evidence: call references, closed-lost post-mortems, and feature-gap flags. Pointers, not transcripts.
• trigger_events: earnings language, hiring signals, executive movement, regulatory filings, competitor signals, and pre-announcement signals.

Column-level contracts live in skills/enterprise-account-based/schema/. The root schema/README.md explains why schema contracts now live per skill. Each schema directory also carries a manifest.json so tests and tools can validate slot contracts without scraping Markdown tables.

Agents

Each agent populates or watches a skill-bound schema slot. One responsibility per agent.

• Signature Authority Miner: populates signature_authority from public filings and contract corpora.
• Persona Graph Builder: assembles persona_graph per account.
• Funnel Telemetry Loader: loads funnel_telemetry from CRM and outreach systems.
• Outcome Telemetry Watcher: watches news, filings, and contract diffs into outcome_telemetry.
• Conversation Evidence Indexer: indexes call references into conversation_evidence.
• Trigger Event Monitor: emits records into trigger_events.
• Anti-Qualification Scorer: computes the consulting/implementation spend ratio using thresholds from the active skill.

See agents/.

Plugins

Plugins are persona-shaped views over the active skill's schema. They do not introduce new data; they assemble role-specific summaries.

• ae: account-executive-shaped views, including target-account dossiers and next-best action scoring.
• sales-leadership: pipeline-level views, board-vs-plan deltas, anti-qualification cohort reporting, and pipeline risk inspection.
• revops: schema health, source quality, and coverage gaps.
• customer-success: renewal risk and expansion-fit radar.
• growth: market-share posture, campaign ROI/payback, search intent, category-demand capture, and intent-to-sequence drafting.
• competitive-intel: competitor signals, battlecards, and permitted public asset-change review.

See plugins/.

Cookbooks

Cookbooks are readable workflows that compose agents and plugins.

• Morning dossier: daily per-account briefing assembled across the reference skill slots.
• Revenue command center: weekly forecast and daily inspection loop that runs schema health, pipeline risk, renewal/expansion radar, and model arbitration.
• Growth command center: category-demand, campaign ROI, and search-intent loop for market creation and capture.
• Intent activation and competitive response: high-intent account routing, compliant sequence drafts, battlecards, and public asset-change review.
• Pre-announcement watcher: planned workflow for publicly observable pre-announcement signals.
• Signal velocity monitor: planned rate-of-change workflow.
• Renewal radar: planned renewal-risk workflow.
• Win/loss interview integrator: planned post-mortem workflow.

See cookbooks/.

Connectors

Connector stubs bind systems of record to the active skill's schema. The minimal code contract lives in connectors/base.py, with an in-memory test connector in connectors/mock.py.

Reference connector names include Salesforce, Gong, Outreach, Slack, Google Drive, 6sense, ZoomInfo, Search Console, GA4, and ad platforms. Connectors default to read-only. Write operations require explicit operator opt-in. Forks may bind different connector names in their own SKILL.md.

See connectors/.

Getting Started

See QUICKSTART.md for prerequisites, validation, and the cold-start interview that produces a per-installation CLAUDE.local.md practice profile with an active_skill.

Validation

The repository has no required third-party Python dependencies. Run:

python -m unittest discover -s tests
python evals/run_evals.py
python tools/inspect_skill.py --json
python examples/forks/plg-self-serve/demo.py

The GitHub Actions workflow in .github/workflows/validate.yml runs the tests, evals, forkability tool checks, and smoke checks for every built agent and plugin demo.

Model Arbitration

core/model_arbitration.py provides a token-aware routing policy for each built workflow. It picks the smallest Claude model tier that satisfies context size, reasoning need, relative cost band, and high-stakes escalation flags. This keeps cheap deterministic checks on a fast model path while preserving an explicit escalation route for board-facing forecast or renewal narratives.

Design rationale and public reference links live in docs/revenue_intelligence_design_notes.md, docs/growth_intelligence_design_notes.md, and docs/intent_activation_design_notes.md.

Roadmap

• Day 1. Repo skeleton: schema slots, agent / plugin / cookbook / connector stubs, README + QUICKSTART + CLAUDE.md scaffold.
• Day 2. Schema column contracts, cold-start profile workflow, and first runnable agent stub.
• Day 3. Morning dossier, AE and sales-leadership scorers, trigger-event monitor, anti-qualification scorer, and deferred connector stubs.
• Day 4. Skills make the harness motion-agnostic; example forks prove that the base can bind different slot sets without hardcoding a motion.
• Day 5. Enterprise revenue, growth, intent activation, competitive response, and forkability tooling make the repo executable end to end.

Contributing

See CONTRIBUTING.md for repository scope, what does and does not belong here, and the pull-request checklist. Operator-specific data belongs in CLAUDE.local.md and other locally ignored files, not in this repository.

Influences And Prior Art

This repository draws on the public discourse of revenue-intelligence product categories, the public canon of sales qualification and methodology, and adjacent buyer-behavior analysis traditions. The current architecture is also influenced by the February 2026 Karpathy framing of a maximally forkable repo whose skills fork it into exotic configurations.

Disclaimers

• Drafts, not decisions. All agent and cookbook outputs are drafts intended for human reviewer judgment. Nothing here claims completeness or accuracy.
• Not advice. Outputs do not constitute legal, financial, or investment advice.
• Web monitoring compliance. Pre-announcement-watcher and any similar feature that observes external sites must be used in compliance with each target site's robots.txt and Terms of Service. No unauthorized scraping is performed or condoned.
• No embedded customer or executive data. This repository contains no customer names, executive names, or proprietary methodology references. Anything operator-specific belongs in the cold-start CLAUDE.local.md practice profile produced locally, not in this repo.