Software Development Skills

This skill consolidates autonomous AI agent loop patterns: the Ralph Wiggum exit-code-2 pattern, our manage-loop (manage-continuation.json + manage-loop-hook.py), grind-loop.sh supervisor, completion detection strategies, safety caps, stall detection, and the tradeoffs between simple loops and sophisticated orchestration. Use when designing or debugging autonomous agent loops, choosing between loop approaches, or implementing safety limits. Do NOT use for choosing which command to run (use ai-coding-agents) or for task management (use linear).

Use when building, reviewing, or applying a Balanced Scorecard for strategy execution and performance management: destination statement, strategy map, perspectives, strategic objectives, strategic readiness, measures/KPIs, targets, initiatives, owners, review cadence, cascading, and learning from performance gaps. Covers financial/stewardship, customer/stakeholder, internal process, and learning/growth or organizational-capacity perspectives, with sustainability/ESG embedded or added as a fifth perspective when strategic, and adaptation for business, nonprofit, public-sector, product, and transformation contexts. Do NOT use for upstream strategy formulation alone (use playing-to-win), quarterly goal-setting alone (use okrs), technical service thresholds (use performance-budgets), portfolio allocation (use bcg-matrix), generic internal/external factor inventory (use swot-tows), probability-weighted option valuation (use expected-value), or activity-level value/cost decomposition (use value-chain-analysis).

Use when uncertainty must be updated after new evidence: start with a prior or base rate, estimate how strongly the evidence fits each hypothesis, update to a posterior, and decide what evidence would change the belief next. Covers priors, base rates, likelihood ratios, posterior odds, calibration, evidence independence, false positives in rare events, and base-rate neglect. Do NOT use for decomposing assumptions to primitives (use first-principles-thinking), tracing downstream consequences (use second-order-thinking), recurring issue clustering (use pattern-recognition), or formal source-to-claim citation (use epistemic-grounding).

Use when analyzing product, brand, business-unit, or investment portfolio allocation with the BCG Growth-Share Matrix: market growth, relative market share, stars, cash cows, question marks, dogs/pets, cash generation, invest/harvest/divest/reposition choices, portfolio balance, and modern limitations. Covers defining portfolio units and markets, choosing comparable metrics, plotting quadrants, interpreting cash-flow logic, challenging share-growth assumptions, and handing off to valuation or strategy methods. Do NOT use for product-market growth paths (use ansoff-matrix), macro-environment scanning (use pestel), industry profit-pressure diagnosis (use porters-five-forces), internal/external factor inventory (use swot-tows), durable moat classification (use seven-powers), or quantified option valuation (use expected-value).

Use when organizing or cleaning a project/repository so every script, command, prompt, doc, generated artifact, compatibility shim, and legacy path has one canonical home under version control. Covers owner selection, repo-boundary verification, generated-vs-authored separation, duplicate implementation cleanup, compatibility-shim registration, delete-to-git-history discipline, timestamped changelog explanations, and path-limited commits in the owning repo. Do NOT use for ordinary branching/rebase/tag strategy alone (use version-control), behavior-preserving code restructuring alone (use refactor), documentation-sync routing alone (use doc-updater), or controlled category taxonomy design alone (use taxonomy-design).

Use when deciding whether to route a task to the fast/cheap tier (Claude Haiku) — transcription, polling, format conversion, structured-output slot-filling, small-diff review, high-volume low-latency work — and where the boundary is that should escalate to Sonnet/Opus. Covers the cost/latency floor, the 200K context ceiling (vs the upper tiers' 1M), the absence of the effort knob, the separate rate-limit pool, and when to drop below Haiku to a script. Do NOT use for ordinary multi-step feature work (use claude-sonnet), the hardest reasoning (use claude-opus), loop design (use autonomous-loop-patterns), or Claude API request syntax (read the claude-api reference).

Use when deciding whether to route a task to Anthropic's frontier reasoning tier (Claude Opus) — architecture, multi-file synthesis, hard debugging, security reasoning, long-horizon agentic planning — and when NOT to (mechanical work belongs on a cheaper tier). Covers the Opus capability profile: 1M-token context, the cost/latency premium, adaptive thinking, the Opus-only effort ceiling (xhigh/max), task budgets, prompt-caching minimums, and high-resolution vision. Do NOT use for picking the balanced implementation tier (use claude-sonnet), the fast/cheap tier (use claude-haiku), designing the loop the model runs inside (use autonomous-loop-patterns), or for Claude API request syntax (read the claude-api reference).

Use when deciding whether to route a task to the balanced implementation tier (Claude Sonnet) — feature work, bug fixes, test writing, multi-step code — as the default lane that is cheaper/faster than the frontier tier and more capable than the fast tier. Covers the cost/quality tradeoff vs Opus and Haiku, the shared 1M context window, effort behavior, and the 1M-context subscription billing caveat. Do NOT use for the hardest reasoning/architecture/security work (use claude-opus), high-volume mechanical or low-latency work (use claude-haiku), loop design (use autonomous-loop-patterns), or Claude API request syntax (read the claude-api reference).

Sweller's Cognitive Load Theory (CLT) for agents writing skill content, designing prompts, building UI/dashboards, and authoring documentation. Working memory holds roughly 4 chunks at a time; CLT classifies load into three types — intrinsic (irreducible task difficulty), extraneous (unnecessary load from poor presentation, ELIMINATE), and germane (load that builds schemas, PROMOTE). Use when writing a SKILL.md body (does this section add extraneous load?), designing prompts (am I asking the model to hold too much at once?), building dashboards (what is the per-screen cognitive budget?), or authoring docs (is intrinsic load minimized via segmentation?). Do NOT use for retrieval and context-loading design (use context-management), prompt engineering tactics (use prompt-craft), or instructional design beyond what grounds the theory.

Designing, running, and interpreting controlled experiments to decide whether a change is genuinely better — online A/B tests, randomized controlled trials, and offline hypothesis tests. Covers hypothesis framing, randomization and unit-of-assignment, sample-size and statistical-power calculation, metric design (primary, guardrail, invariant, and the overall evaluation criterion), variance reduction (CUPED/CUPAC) for sensitivity, significance testing and confidence intervals, sequential and anytime-valid testing for the peeking problem, multiple-comparison correction, the trust checklist led by sample-ratio-mismatch, interference-aware designs (cluster and switchback), long-term holdouts and reverse experiments, bandit/adaptive-allocation boundaries, experiment ethics and blast-radius limits, quasi-experimental fallbacks when randomization is infeasible, and the validity threats that quietly invalidate results (novelty and primacy effects, interference/SUTVA violations, Simpson's paradox, and survivorship).

Use when deciding whether to route a task to Google's Gemini Pro frontier model (current Gemini 3.1 Pro generation, moving to Gemini 3.5 Pro) instead of Claude Opus or GPT-5 — especially for very-large-context reasoning (1M-token input), whole-codebase or long-document analysis, native multimodal input (audio/video/PDF/image in one call), or capability-per-dollar at the frontier. Covers the context window, the 200K context-tier pricing cliff, multimodal support, and the per-lane comparison against Claude Opus / GPT-5. Do NOT use for choosing the cheap/fast tier (use gemini-flash), for general agent-system architecture (use agent-engineering), or for dispatching among local skills (use skill-router).

Use when deciding whether to route a task to OpenAI's GPT-5.5 frontier model versus Claude Opus or Sonnet — picking the model lane for infrastructure scripts, CI pipelines, concrete implementation, analytical code review, security review, or CLI/terminal-heavy work, and weighing GPT-5.5's context window, pricing, and per-benchmark strengths against the Claude tiers for the same task. Covers the decision-useful capability and pricing facts, and the boundary against the Claude routing skills. Do NOT use for running the GPT model through a harness (use codex), for choosing among the Claude tiers themselves (use claude-opus / claude-sonnet), or for routing among local skills at request time (use skill-router).

Solve problems by thinking backward - what would cause failure?

Use when diagnosing organizational alignment with the McKinsey 7S framework: shared values, strategy, structure, systems, style, staff, and skills; how the seven elements reinforce or conflict; change-readiness gaps; and implementation risks. Covers internal organization effectiveness, operating-model alignment, transformation diagnostics, post-merger integration checks, and strategy-to-execution gap analysis. Do NOT use for external industry profit-pressure diagnosis (use porters-five-forces), activity-level value and cost mapping (use value-chain-analysis), durable resource advantage testing (use vrio), broad internal/external option generation (use swot-tows), or execution goal-setting alone (use okrs).

Enforces disciplined, complete, step-by-step execution that prevents the LLM failure modes of scope reduction, finding filtering, step skipping, assumed verification, and sycophantic output compression. Provides the explanatory model for WHY agents fail at completeness (RLHF training rewards shorter, cleaner, more positive outputs — 58% sycophancy rate measured in frontier models) and the structural countermeasures: pre-task declarations, step-level evidence receipts, generation/criticism separation, explicit completeness claims, and anti-pattern detection. Use when executing audits, producing reports, creating Linear tasks from findings, verifying acceptance criteria, or any task where completeness and honesty matter more than brevity. Do NOT use for task workflow sequencing (use task-execution), quality definitions per artifact type (use quality-doctrine), or the generate-critique-revise loop mechanics (use self-review-pattern).

Enforces five non-negotiable quality pillars that counter the shortcuts an agent takes under training pressure: complete reporting (show ALL items, never filter unilaterally), verification (no claim of works/done/exists without a tool-call receipt in the same turn), thoroughness (every acceptance criterion individually verified; docs ship with the change), enrichment ('improve' means add capability, never trim), and anti-shortcut (exhaust deterministic lookup before guessing; findings demand action, not just filing). Use when reviewing any enumerated output for completeness, when an agent claims something works without evidence, when marking a task done, when asked to 'improve' or 'clean up' anything, or when findings are filed without being acted on. Do NOT use for the deep explanatory model of WHY completeness fails or the step-level execution mechanics (use methodical), for the cross-domain quality-standards catalog like OWASP/WCAG/SOLID (use best-practice), for scoring whether a result is good enough (use evaluation), or for compression after a complete enumeration exists (use summarization).

Objectives and Key Results goal-setting methodology for turning strategy, quarterly priorities, product goals, or team focus areas into outcome-oriented Objectives, measurable Key Results, review cadences, and learning loops.

Use when deciding WHICH free or cheap agent model to route a piece of work to (e.g. MiniMax M3 Free, NVIDIA Nemotron, GLM, GPT-5 Nano, or another OpenCode Zen free-tier model), what each is good at, where its quality ceiling sits, and when to escalate the same work to a frontier model instead. Covers cost-routing: matching deterministic/mechanical/high-volume work to the cheapest model that clears its bar, and the hard rule that quality-creating and quality-judging work never goes to a free/cheap model. Do NOT use for choosing or operating the OpenCode runtime itself (use opencode), for authoring the agent loop (use autonomous-loop-patterns), or for GitHub Copilot premium-request budgeting (use github-copilot).

Use when scanning an external macro environment with PESTEL/PESTLE (and its variant family STEEPLE, STEEPLED, PESTLIED, STEEP, DESTEP, LoNGPESTLE): political, economic, social, technological, environmental, and legal forces; evidence quality and recency; geography/jurisdiction and local/national/global level; time horizon; uncertainty; weak signals; impact/probability scoring; factor interactions; bias checks; opportunity/threat implications; assumptions; action conversion; and monitoring triggers. Covers external-environment scanning before strategy choices, market-entry reviews, strategic planning, product/service context, policy-aware planning, and risk/opportunity surfacing. Do NOT use for internal capability diagnosis (use swot-tows or a capability method), industry profit-pressure diagnosis (use porters-five-forces), value-curve redesign (use blue-ocean-strategy), integrated strategy cascades (use playing-to-win), product positioning (use positioning), or quantified option comparison (use expected-value).

Use when turning a vague business strategy, product strategy, market-entry decision, or initiative plan into an integrated Playing to Win strategy cascade: winning aspiration, where to play, how to win, must-have capabilities, and management systems. Covers Lafley/Martin choice-making, fit across the five choices, trade-off pressure, reverse tests, capability-system alignment, and the difference between strategy and planning. Do NOT use for competitive-industry structure analysis (use a five-forces skill when available), generic backlog scoring (use prioritization), or broad process-gate design (use methodology).

Use when preparing, structuring, or pressure-testing a negotiation with principled negotiation / BATNA: separate people from the problem, focus on interests rather than positions, invent options for mutual gain, use objective criteria, define BATNA / reservation value / ZOPA, and decide whether to accept, improve, or walk away. Do NOT use for generic persuasion copy (use microcopy or writing-humanizer), unilateral conflict diagnosis without a deal table, competitive strategy analysis (use playing-to-win, porters-five-forces, or seven-powers), or broad decision scoring without counterparties (use expected-value or prioritization).

Use when facing a new problem and unsure which problem-solving methodology or foundational skill to apply first. Routes between first-principles-thinking, pattern-recognition, mental-models, constraint-awareness, and task-analysis by classifying the problem type. Activate before choosing any other foundational skill. Do NOT use to execute the selected approach (use the specific routed skill directly), for implementation work (use the relevant engineering skill), or when the correct approach is already known.

Use when a decision, intervention, policy, product change, metric target, or system fix needs analysis beyond the immediate effect: ask what happens next, how actors respond, what incentives change, and what delayed or compounding consequences follow. Covers consequence chains, time horizons, stakeholder reactions, feedback loops, second- and third-order effects, and guardrail design. Do NOT use for reversing a goal to find failure paths (use inversion), decomposing assumptions to primitives (use first-principles-thinking), recurring issue clustering (use pattern-recognition), or system leverage-level diagnosis (use meadows-leverage-points when available).

This skill records the formal ontological commitments of the skill-graph protocol — what entities exist (Skill, ConceptCard, Eval, ComprehensionEval, Relation, Manifest, Audit, Run, Verdict, DriftBaseline, TruthSource), what identity criteria distinguish them, and what axioms govern their state and relationships (e.g., a Skill has exactly one schemaversion; a Verdict tier ordering is APPLICABLE > PROVISIONAL > UNVERIFIED; a ComprehensionEval must carry a dimension tag from the 7 canonical dimensions). Use when proposing a new field or entity in the protocol, designing a schema migration, debating whether a new concept deserves its own protocol entity vs. an attribute on an existing one, or debugging cross-entity inconsistencies in the audit pipeline.