feat: Add persona-based agents (Startup CTO, Growth Marketer, Solo Founder)

[alirezarezvani]

Summary

Adds persona-based agents — a complementary layer on top of existing agents and skills.

What's a Persona?

An agent definition that goes beyond task execution. Each persona includes:

• Identity & mindset — how they think, what they prioritize
• Curated skills — pre-loaded skill set for their domain
• Workflows — multi-step processes (MVP Sprint, Launch Sequence, etc.)
• Communication style — personality-driven, not generic
• Handoffs — when to escalate to another persona/agent

3 Starter Personas

Persona	Domain	Skills	Workflows
startup-cto	Engineering + Strategy	9 skills	Tech Stack Selection, Architecture Review, Technical Due Diligence, Incident Response
growth-marketer	Marketing + Growth	9 skills	90-Day Content Engine, Product Launch, Conversion Audit
solo-founder	Cross-domain	10 skills	MVP in 2 Weeks, Weekly Sprint (Solo), Should I Build This?, Pricing Decision

Files Added

agents/personas/
├── README.md           # What personas are + how to use
├── TEMPLATE.md         # Format spec for creating new personas
├── startup-cto.md      # Technical co-founder
├── growth-marketer.md  # Bootstrapped growth
└── solo-founder.md     # Indie hacker all-in-one

Coexistence with Existing Agents

• Agents (agents/) → focused task execution
• Personas (agents/personas/) → role embodiment with curated skills + workflows
• Both formats supported. No existing files modified.

Multi-tool Support

Personas convert to all supported tools via existing scripts/convert.sh and scripts/install.sh.

---

[claude]

PR Review: feat: Add persona-based agents

Great concept and well-structured implementation overall. The persona pattern fills a real gap between task-focused agents and role-based collaboration. A few issues need to be addressed before merging.

---

Broken Skill References

Two skill paths referenced in the personas do not exist in the repository:

engineering/architecture-pattern-selector — referenced in all 3 personas (startup-cto, growth-marketer via secondary, solo-founder). The engineering/ directory contains no such skill. The closest thing is likely engineering-team/senior-architect. Either create the skill or correct the reference.

engineering/cost-estimator — referenced in startup-cto.md (YAML + Secondary) and solo-founder.md (YAML + Secondary). Also absent from the engineering/ directory. No matching skill exists.

These broken references will silently fail when users try to activate the referenced skills.

---

Mismatched Command Names

The following commands referenced in YAML frontmatter don't match any existing command files in commands/:

Persona	Referenced Command	Status
startup-cto	/tech-debt-audit	Exists as /tech-debt (different name)
startup-cto	/architecture-review	Does not exist
growth-marketer	/content-plan	Does not exist
growth-marketer	/launch-checklist	Does not exist
solo-founder	/launch-checklist	Does not exist
solo-founder	/sprint-plan	Exists ✓

Commands that don't exist will produce confusing behavior. Either create the missing commands or update the frontmatter to reference existing ones (e.g., /tech-debt-audit → /tech-debt).

---

Skill Reference Inconsistency in startup-cto.md

engineering-team/senior-architect appears in the YAML frontmatter skills list but is not mentioned anywhere in the markdown Skills section (not in Primary or Secondary). This is likely an oversight — either add it to Secondary with a trigger condition or remove it from the frontmatter.

---

Missing Documentation Updates

• agents/CLAUDE.md — No updates to document the new persona format/structure for contributors creating new personas.
• Root CLAUDE.md — Still lists "16 agents" in the count. If personas count toward agents, this should be updated; if they're a separate concept, add a note.
• CHANGELOG.md — No changelog entry for this addition.

---

Minor Observations

Handoff naming inconsistency: Handoff targets mix un-prefixed persona names (startup-cto, growth-marketer) with cs--prefixed agent names (cs-financial-analyst, cs-agile-product-owner, cs-ux-researcher). This is correct by design (personas vs agents) but the README could mention this distinction explicitly so users don't get confused about why some handoffs have cs- and others don't.

No personas index in agents/: The existing agents directory has no index file. Adding an entry in agents/CLAUDE.md or agents/README.md (if one exists) would make personas discoverable without requiring users to know about the personas/ subdirectory.

---

What Works Well

• Consistent and complete YAML frontmatter across all 3 persona files
• Clear separation of Primary vs Secondary skills with trigger conditions — this is a good design pattern
• Anti-patterns sections are genuinely useful and not generic
• Handoffs table is well-designed and actionable
• TEMPLATE.md enables contributors to create new personas without guessing the format
• README's Agents vs Personas comparison table is clear and concise
• scripts/convert.sh and scripts/install.sh exist and support the multi-tool claim in the README

---

Summary

The concept is solid and the content quality is high. The main blockers are the broken skill references (engineering/architecture-pattern-selector and engineering/cost-estimator) and the missing commands. Fix those and update the documentation references, and this is ready to merge.

[devin-ai-integration]

Devin Review found 7 new potential issues.

View 5 additional findings in Devin Review.

[claude]

Skills Audit Report

Date: 2026-02-15
Auditor: Automated Skill Quality Audit
Scope: Recently added skills in business-growth/, finance/, marketing-skill/campaign-analytics/, project-management/

---

Executive Summary

The recently added skills fall into two distinct tiers:

1. Business-growth, Finance, and Campaign Analytics skills — Genuinely impressive. Production-ready Python tooling, deep domain frameworks, real structured outputs. These would make a domain practitioner say "this actually knows what it's doing."

2. Project Management skills — A mixed bag. The Atlassian-specific skills (jira-expert, confluence-expert, atlassian-admin, atlassian-templates) have strong knowledge-base content. The scrum-master and senior-pm skills are thin and generic. None of the PM skills have scripts or assets — they're pure prompt-engineering skills, which is a fundamentally different (and weaker) category.

Overall: 4 POWERFUL, 1 SOLID, 4 SOLID, 2 GENERIC, 1 WEAK

---

Detailed Skill Audits

---

1. business-growth/customer-success-manager

Code Quality: EXCELLENT

• 3 Python scripts (438 + 487 + 414 = 1,339 lines total)
• Well-structured: proper typing, argparse CLI, JSON/text dual output, error handling
• Zero external dependencies (stdlib only) — deliberate, documented design choice
• health_score_calculator.py: Multi-dimensional weighted scoring with segment-aware benchmarks (Enterprise/Mid-Market/SMB). Not placeholder math — real configurable thresholds, normalization logic, trend analysis
• churn_risk_analyzer.py: Behavioral signal detection with renewal urgency multipliers
• expansion_opportunity_scorer.py: Whitespace mapping and effort-vs-impact prioritization
• All scripts actually runnable with provided sample data

Problem-Solving Quality: EXCELLENT

• Health scoring framework reference (80+ lines) explains why each dimension is weighted as it is — genuinely pedagogical
• Real CS playbooks: not "be proactive" platitudes but specific intervention triggers (e.g., "if health score drops below yellow for 2 consecutive periods, escalate")
• QBR template is production-ready — has ROI summary tables, value-delivered sections, next-quarter planning
• Success plan template, onboarding checklist, executive business review — all structured with fill-in tables
• Uses real industry frameworks: DAU/MAU ratio, NPS scoring methodology, multi-threading depth

Structure: STRONG

• SKILL.md has proper frontmatter, TOC, input/output schemas, limitations section
• References are actually used by the scripts (health-scoring-framework.md maps directly to score calculation logic)
• Assets include sample data AND expected output JSON for validation

Verdict: POWERFUL ⭐
Evidence: A CS leader could hand this to a team and they'd have a working health scoring system same day. The weighted scoring model with segment-aware thresholds is exactly how real CS platforms (Gainsight, Totango) work. The scripts produce structured JSON that could feed a dashboard.

---

2. business-growth/revenue-operations

Code Quality: EXCELLENT

• 3 scripts (496 + 531 + 658 = 1,685 lines total) — the largest script set
• pipeline_analyzer.py: Coverage ratios, stage conversion rates, sales velocity formula (Opportunities × Avg Deal × Win Rate / Cycle), deal aging detection, concentration risk warnings
• forecast_accuracy_tracker.py: MAPE calculation, period-over-period accuracy trending
• gtm_efficiency_calculator.py: CAC, LTV, CAC payback period, magic number, burn multiple — these are real SaaS metrics, not made up
• Proper CLI args, dual output format, input validation

Problem-Solving Quality: EXCELLENT

• RevOps metrics guide references real benchmarks (3-4x pipeline coverage, magic number >0.75)
• Pipeline management framework covers qualification methodology
• GTM efficiency benchmarks are industry-standard (Bessemer, OpenView style)
• Templates: pipeline review, forecast report, GTM dashboard — all structured with metric tables

Structure: STRONG

• Consistent with customer-success-manager pattern
• Sample data files for all three scripts
• Expected output JSON for validation

Verdict: POWERFUL ⭐
Evidence: The pipeline analyzer alone replaces basic Salesforce reporting. The GTM efficiency calculator uses the exact metrics VCs and board members ask for (magic number, burn multiple, CAC payback). A RevOps manager would find real utility here.

---

3. business-growth/sales-engineer

Code Quality: EXCELLENT

• 3 scripts (557 + 525 + 765 = 1,847 lines total) — largest individual script set
• rfp_response_analyzer.py: Weighted coverage scoring (Full/Partial/Planned/Gap × Must-have/Should-have/Nice-to-have), automated bid/no-bid recommendation with configurable thresholds
• competitive_matrix_builder.py: Feature-by-feature comparison with differentiator/vulnerability identification
• poc_planner.py: Timeline generation, resource planning, success criteria definition, evaluation scorecards
• 765-line POC planner is genuinely comprehensive

Problem-Solving Quality: EXCELLENT

• 5-phase workflow (Discovery → Solution Design → Demo → POC → Close) maps to real SE methodology
• RFP analyzer produces structured gap analysis with mitigation strategies — not just "you have gaps"
• Competitive positioning framework includes feature-level comparison, not just "we're better"
• Demo script template and POC scorecard are practitioner-level artifacts
• Technical proposal template has architecture section

Structure: STRONG

• Same consistent pattern as other business-growth skills
• Rich asset set: demo script template, POC scorecard, technical proposal template, sample RFP data
• References cover competitive positioning, POC best practices, RFP response methodology

Verdict: POWERFUL ⭐
Evidence: The RFP analyzer with weighted coverage scoring and bid/no-bid recommendation is something SEs actually need and usually do in spreadsheets. The POC planner at 765 lines is the most substantive single script in this batch. A pre-sales team could adopt this immediately.

---

4. finance/financial-analyst

Code Quality: EXCELLENT

• 4 scripts (432 + 449 + 406 + 494 = 1,781 lines total)
• ratio_calculator.py: 20+ ratios across 5 categories (profitability, liquidity, leverage, efficiency, valuation) — ROE, ROA, DSCR, DSO, EV/EBITDA, PEG ratio
• dcf_valuation.py: Full DCF model with WACC via CAPM, 5-year projections, terminal value (perpetuity growth AND exit multiple methods), two-way sensitivity analysis, equity bridge
• budget_variance_analyzer.py: Favorable/unfavorable classification by department and category
• forecast_builder.py: Driver-based forecasting with scenario modeling (base/bull/bear)
• All stdlib only, handles edge cases (inf values in JSON serialization)

Problem-Solving Quality: EXCELLENT

• DCF model implements real finance: CAPM cost of equity, after-tax cost of debt, terminal value via both methods, sensitivity matrix — this is textbook corporate finance done correctly
• Ratio guide includes interpretation context (not just "here's the number" but "here's what it means")
• Valuation methodology reference explains when to use DCF vs. comparables vs. precedent transactions
• Forecasting best practices cover driver-based vs. trend-based approaches
• Variance report template is exactly what FP&A teams produce monthly

Structure: STRONG

• Consistent format with other skills
• 4 scripts (most of any skill) — comprehensive coverage of analyst workflow
• Sample data, expected output, 3 templates (DCF, forecast, variance)

Verdict: POWERFUL ⭐
Evidence: The DCF valuation model alone is genuinely useful — it implements WACC calculation, cash flow projection, terminal value via two methods, and sensitivity analysis. A junior analyst could use this as a learning tool; a senior analyst could use it for quick-and-dirty valuations. The sensitivity table output is exactly what you'd see in an investment banking pitch book.

---

5. marketing-skill/campaign-analytics

Code Quality: VERY GOOD

• 3 scripts (347 + 459 + 305 = 1,111 lines total) — smallest script set but still substantive
• attribution_analyzer.py: 5 attribution models (first-touch, last-touch, linear, time-decay, position-based) — these are the real standard models used in marketing analytics
• campaign_roi_calculator.py: ROI, ROAS, CPA, CPL, CAC with industry benchmarking
• funnel_analyzer.py: Stage-by-stage conversion rates, drop-off identification, bottleneck detection
• Time-decay model uses configurable half-life parameter — not just a label

Problem-Solving Quality: VERY GOOD

• Attribution models guide explains when to use each model (rare — most resources just list them)
• Funnel optimization framework covers real concepts (stage-specific interventions, not just "improve conversion")
• Campaign metrics benchmarks provide industry reference points
• A/B test template and channel comparison template are useful artifacts

Structure: STRONG

• Consistent with business-growth pattern
• References tied to script functionality
• Sample data with customer journeys for attribution testing

Verdict: SOLID (borderline POWERFUL)
Evidence: The 5 attribution models are correctly implemented and genuinely useful for any marketing team not yet using a dedicated attribution platform. However, the funnel analyzer (305 lines) is thinner than the equivalent scripts in other skills, and the overall scope is narrower than the business-growth skills.

---

6. project-management/jira-expert

Code Quality: N/A — No scripts

Problem-Solving Quality: GOOD

• JQL examples reference is genuinely useful — covers sprint queries, team workload, SLA tracking, change management queries
• Automation examples reference covers real Jira automation rules
• SKILL.md has comprehensive workflow descriptions for project creation, workflow design, JQL building
• Actually teaches JQL syntax with practical examples, not just theory

Structure: ADEQUATE

• No scripts, no assets, no sample data
• But the references are substantive (415 + 423 = 838 lines of reference material)
• Workflows reference other PM skills (Scrum Master, Confluence Expert) — good cross-linking

Verdict: SOLID
Evidence: The JQL examples alone are a legitimate reference resource. The automation examples cover real-world rules. But without scripts or structured output tooling, this is fundamentally a knowledge-base skill, not a tool skill. It makes Claude better at Jira advice but doesn't produce artifacts.

---

7. project-management/confluence-expert

Code Quality: N/A — No scripts

Problem-Solving Quality: GOOD

• Templates reference (725 lines) contains 10+ ready-to-use Confluence page templates: meeting notes, decision log, project status, runbook, postmortem, ADR, onboarding guide
• Space architecture guidance is practical and specific (max 3 levels deep, naming conventions)
• Macro usage examples are helpful for Confluence power users

Structure: ADEQUATE

• Strong reference content compensates for lack of scripts
• Templates are the actual artifact output — when Claude uses this skill, it produces Confluence pages

Verdict: SOLID
Evidence: The templates reference is the real value here — it's a curated library of production-quality Confluence page templates. A team setting up Confluence from scratch would find this genuinely useful. The space architecture guidance reflects real organizational experience.

---

8. project-management/atlassian-admin

Code Quality: N/A — No scripts

Problem-Solving Quality: GOOD

• SKILL.md is comprehensive at 414 lines covering user provisioning, deprovisioning, group management, permission schemes, security configuration
• Workflows are procedural and actionable (step-by-step with handoffs to other skills)
• Permission scheme design section is practical — distinguishes public/team/restricted/confidential project types
• SSO/SAML and security policy coverage is relevant

Structure: ADEQUATE

• No references, no assets — all content in SKILL.md
• Good cross-references to other PM skills (Jira Expert, Confluence Expert)

Verdict: SOLID
Evidence: The user provisioning/deprovisioning workflows with audit steps reflect real admin concerns (content reassignment before account deletion). Permission scheme design is specific enough to be useful. But without reference docs or scripts, it's a well-written playbook rather than a tool.

---

9. project-management/atlassian-templates

Code Quality: N/A — No scripts

Problem-Solving Quality: GOOD

• SKILL.md at 751 lines is the longest PM skill — contains actual template content inline
• Template creation process (10-step) and modification process (8-step) are well-structured
• Contains ready-to-use templates: meeting notes, decision log, sprint planning, retrospective, project charter
• Blueprint development workflow is practical

Structure: ADEQUATE

• All content in SKILL.md — no separate references or assets
• Templates are embedded directly rather than in a templates/ directory

Verdict: SOLID
Evidence: The templates themselves are the deliverable, and they're decent. The template governance process (versioning, deprecation, migration) shows organizational maturity. However, significant overlap with confluence-expert/references/templates.md raises questions about redundancy.

---

10. project-management/scrum-master

Code Quality: N/A — No scripts

Problem-Solving Quality: MEDIOCRE

• SKILL.md at 189 lines is thin — covers basic Scrum ceremonies at a surface level
• Nothing here goes beyond what's in the Scrum Guide
• No velocity tracking formulas, no capacity planning models, no sprint health metrics
• Retro formats reference (336 lines) is the saving grace — covers Start/Stop/Continue, Glad/Sad/Mad, 4Ls, Sailboat, DAKI formats with actual process steps

Structure: WEAK

• No assets, no sample data
• Single reference file
• Cross-references to Jira Expert and Confluence Expert add some value

Verdict: GENERIC
Evidence: A certified Scrum Master would find nothing new here. The retro formats reference is genuinely useful but is the only substantive content. The SKILL.md reads like a job description, not a methodology. No metrics, no anti-patterns, no "when sprints go wrong" playbooks. Missing: burndown analysis tools, velocity prediction, capacity planning scripts.

---

11. project-management/senior-pm

Code Quality: N/A — No scripts

Problem-Solving Quality: WEAK

• SKILL.md at 146 lines is the thinnest skill in the entire batch
• references/api_reference.md is literally a placeholder: "This is a placeholder for detailed reference documentation. Replace with actual reference content or delete if not needed."
• Content is generic PM advice: "develop product roadmaps aligned with business objectives," "identify and mitigate project risks"
• No frameworks, no decision models, no risk quantification methods
• No RACI template, no project charter template despite mentioning them

Structure: WEAK

• Placeholder reference file is a red flag
• No assets, no templates, no sample data
• Mentions creating artifacts (RACI matrix, project charter) but provides no templates

Verdict: WEAK
Evidence: The placeholder reference file tells the whole story — this skill was scaffolded but never completed. A senior PM would find nothing actionable. Compare to the financial-analyst skill (1,781 lines of working code + templates) vs. this (146 lines of generic advice + a placeholder). This is "act as a Senior PM" prompting dressed up as a skill.

---

Comparative Analysis

Skill	Scripts (LOC)	References	Assets/Templates	Verdict
customer-success-manager	3 (1,339)	3 deep	5 templates + sample data	POWERFUL
revenue-operations	3 (1,685)	3 deep	7 templates + sample data	POWERFUL
sales-engineer	3 (1,847)	3 deep	5 templates + sample data	POWERFUL
financial-analyst	4 (1,781)	3 deep	4 templates + sample data	POWERFUL
campaign-analytics	3 (1,111)	3 deep	5 templates + sample data	SOLID
jira-expert	0	2 substantive	0	SOLID
confluence-expert	0	1 (725 lines)	0	SOLID
atlassian-admin	0	0	0	SOLID
atlassian-templates	0	0	0	SOLID
scrum-master	0	1 (336 lines)	0	GENERIC
senior-pm	0	1 (placeholder!)	0	WEAK

Key Observations

What Works (business-growth, finance, campaign-analytics)

1. Scripts that actually compute things — Not wrappers, not boilerplate. Real algorithms with real business logic (DCF valuation, attribution modeling, health scoring)
2. Zero external dependencies — Deliberate stdlib-only design means they run anywhere, immediately
3. Dual output format — JSON for automation, text for humans. This is good engineering
4. Sample data + expected output — Enables validation and serves as documentation
5. References that explain why — The health scoring framework doesn't just list metrics; it explains why each dimension is weighted as it is
6. Templates that are fill-in-ready — QBR template, variance report, demo script — these save real time

What Doesn't Work (parts of project-management)

1. Senior PM is unfinished — Placeholder reference file, no templates despite claiming to produce them
2. Scrum Master is generic — Doesn't exceed the Scrum Guide in depth or utility
3. No scripts in any PM skill — The business-growth skills prove that scripts add massive value. The PM skills could have had: sprint velocity calculator, capacity planner, risk matrix scorer, RACI generator
4. Two-tier quality — The gap between POWERFUL and WEAK skills in the same repo is jarring

Recommendations

1. Senior PM needs a complete rewrite or removal — The placeholder reference is unacceptable. Either build it to the standard of financial-analyst (scripts + real frameworks) or don't ship it
2. Scrum Master needs depth — Add velocity tracking scripts, burndown analysis, capacity planning calculator, sprint health scorer
3. PM skills should get scripts — Even simple ones: RACI matrix generator, risk register scorer, project status report formatter
4. Deduplicate PM templates — atlassian-templates and confluence-expert overlap significantly
5. Add expected_output.json to PM skills — If they can't have scripts, at least define what "good output" looks like

---

Report generated 2026-02-15. Skills assessed against the bar: "Would this make someone say 'holy shit, this actually knows what it's doing?'"

Business-growth and finance skills clear that bar. Campaign-analytics nearly does. PM skills mostly don't.

[claude]

PR Review test post

[claude]

PR Review: Persona-Based Agents

[claude]

Great concept. The distinction between task agents (focused execution) and personas (role embodiment) is well-articulated and fills a real gap. The persona content is high quality: detailed workflows, concrete communication examples, and actionable success metrics. The non-destructive approach (no existing files modified) is appreciated.

That said, there are several structural issues against established conventions that should be addressed before merge.

[claude]

Issues

1. YAML Frontmatter Deviates from Required Schema

Per agents/CLAUDE.md, every agent file must include: name, description, skills, domain, model, and tools (as an array like [Read, Write, Bash, Grep, Glob]). The personas are missing skills, domain, and model, and introduce undocumented fields (color, emoji, vibe). The tools field is also a string instead of an array. If personas intentionally use a different schema, this deviation needs to be documented in agents/CLAUDE.md.

2. Broken Script Paths in README

README.md references ./scripts/convert.sh and ./scripts/install.sh. From agents/personas/, the correct relative paths are ../../scripts/convert.sh and ../../scripts/install.sh. The current paths will fail for anyone following the README instructions.

3. No Skill Integration

Existing agents link to specific skills via relative paths (e.g. ../../engineering/database-designer) and reference Python automation tools directly. The personas list skills conceptually but do not reference actual paths or tools, meaning they cannot leverage the 254 automation tools in the library. The README should explicitly acknowledge this as an intentional design difference from task agents, not an omission.

4. Agent Catalog Not Updated

The production agents catalog in agents/CLAUDE.md lists 16 agents and should be updated to include these 3 new personas. The main CLAUDE.md agent count would also need updating if these are treated as first-class agents.

5. No cs- Prefix -- Convention Undocumented

agents/CLAUDE.md requires the cs-* prefix for all agents. The personas skip this, which may be intentional, but there is no documentation explaining why. TEMPLATE.md should note this explicitly so future persona authors are not confused.

6. Missing Related Agents Cross-References

The agent quality checklist requires a Related Agents section. startup-cto should link to cs-cto-advisor, growth-marketer to cs-content-creator, solo-founder to both, etc.

[claude]

Suggestions

• Either add the missing YAML fields (skills, domain, model) or document the persona schema as a new, separate format in agents/CLAUDE.md
• Fix script paths in README.md to use ../../scripts/convert.sh
• Add a Relationship to Skills section to each persona explaining conceptual skill mapping
• Update the agent catalog in agents/CLAUDE.md with the 3 new personas
• Add a naming convention note to TEMPLATE.md explaining why personas do not use the cs- prefix

Positive Notes

• The concept is genuinely additive -- personas occupy a meaningfully different use case than task agents
• Workflow documentation (MVP in 2 Weeks, Incident Response, Conversion Audit, etc.) is concrete and directly useful
• TEMPLATE.md is a solid foundation for future persona authors
• Communication style examples with actual sample quotes are excellent and make personas feel real
• The Coexistence with Existing Agents framing in the PR description is the right mental model and should be surfaced in the README itself

Content quality is high. The issues are mostly conformance gaps that should be straightforward to address.