Skills

A collection of reusable AI agent skills — structured frameworks that guide agents to reason and act more rigorously.

---

Skills

hypothesis-driven-task-execution

A scientific thinking framework that treats every piece of information — user requests, agent responses, feedback, and experiment results — as a hypothesis to be verified through falsification-first experimentation.

When to use: investigation, debugging, root cause analysis, validating assumptions, or any task involving uncertainty or competing explanations.

Key concepts:

• 5 Hypothesis Layers (L1–L5): source-aware classification of every claim (user request → agent response → user feedback → experiment interpretation → external knowledge)
• 6-Phase Protocol: Observe → Hypothesize → Predict → Experiment → Analyze → Conclude
• Falsification-first: attempt to disprove a hypothesis before seeking confirmation
• Confidence scale: speculation → low → medium → high (based on independent falsification attempts survived)
• Safeguards: hypothesis budget (max 5 active), iteration cap (max 5 primary loops), confirmation-bias guards

git-notes-agent-memory

Cross-session agent memory stored as Git notes: a canonical handoff note (refs/notes/agent-memory) plus optional per-agent notes (refs/notes/am-agents/<id>) with ./agent-memory.sh aggregate for overlap heuristics (wasteful_duplicate / healthy_parallel). Ships agent-memory.sh (bash only) and SCHEMA.md.

When to use: session handoffs, multi-agent coordination, resuming after context loss, or parallel work where you want structured overlap detection without using raw Git merge conflicts as the primary signal.

Key concepts:

• Two modes: canonical cumulative note vs per-agent refs + aggregation (see SCHEMA.md Optional — Phase 3 fields: task_id, hypothesis_hash)
• Refs: am-agents is a sibling of agent-memory (Git cannot nest agent-memory/agents/... under the leaf ref agent-memory)
• Experiments: reproducible scenarios in the agent-memory-lab repo (EXPERIMENT_PHASE2.md, EXPERIMENT_PHASE3.md)

api-goal-canvas-designer

Builds an API Goal Canvas (Whos | Whats | Hows | Inputs | Outputs | Goals) from rough API, product, or workflow descriptions. Each row is a coherent Who → What → How → Input → Output → Goal flow record, not a loose collection of cells. Detailed rules, heuristics, and templates live in references/REFERENCE.md.

When to use: designing or reviewing APIs, restructuring feature-centric requirements into actor-goal flows, or detecting missing producers/consumers, orphan records, and over-segmented request flows.

Key concepts:

• Six-column canvas: request-level meaning over endpoint-level labels
• Connectivity checks: missing producer / missing consumer / orphan record detection
• Decomposition rules: break vague verbs (manage, process, integrate) into concrete actions; separate direct users from indirect consumers, internal services, external systems, operators, and batch actors
• Tags: [confirmed], [inferred], [missing], [orphan]
• Completion states: INCOMPLETE → DRAFT_COMPLETE → FINAL_COMPLETE, paired with Discovery / Draft / Final response modes

---

Using both skills together (synergy)

Hypothesis-driven task execution gives you how to run an investigation: explicit falsifiable claims, predictions, experiments, and a hypothesis ledger with confidence levels.

Git notes agent memory gives you where that state survives: durable notes on commits, readable by any bash-capable agent after a restart or handoff.

Together they support:

1. Durable scientific runs — Record the current hypothesis ledger, next falsification experiment, and open questions in a canonical note after each commit; the next session reads the same structure the hypothesis skill prescribes instead of starting from a blank slate.

2. Multi-agent experiments — Parallel agents write to refs/notes/am-agents/<id> with task_id and hypothesis_hash aligned to your hypothesis labels; aggregate surfaces wasteful duplicate work (same task + same hypothesis variant) vs healthy parallel exploration (same task + different hypothesis variants) — a concrete bridge between “rival hypotheses” in the framework and observable overlap in the repo.

3. Evidence that outlives the chat — Falsification outcomes and ledger updates can live in note bodies (## Context, decisions) so “what we tried and what survived” stays tied to the git timeline, not only to model context.

Load both skills when you are running long-running or multi-session investigations that should stay structured (hypothesis protocol) and reproducible (git-addressable memory).

---

Usage

Each skill is defined in a SKILL.md file and can be referenced in agent system prompts, task instructions, or tool configurations. Skills are composable — apply full rigor for complex investigations, lighter application for mechanical tasks.