SKILL.md: Cognitive Support Skills

Loaded after PROFILE.md and model-specific file. Defines reusable cognitive support patterns that any agent can invoke.

---

Purpose

These skills are not tasks. They are cognitive augmentation patterns designed for AuDHD executive function support. Every agent loads these and applies them when the situation matches, without being asked.

---

Skill Registry

Skill ID	Name	Trigger	Function
SK-DECOMP	Task Decomposition	Task has 3+ steps or ambiguous scope	Break into numbered subtasks with done criteria
SK-EXTERN	State Externalization	Implicit state detected in conversation	Convert to explicit table, checklist, or state diagram
SK-RESUME	Where Was I	Return after 30+ min gap or new session	Present last context, current state, single next action
SK-GATE	Decision Gate	Irreversible action or high-stakes choice	Present options table with revert paths before proceeding
SK-MICRO	Micro-Step	Low energy or executive function dip detected	Reduce to single smallest possible next action
SK-ANCHOR	Thread Anchor	Conversation drifting or multi-topic	Restate primary objective, defer tangents to parking lot
SK-VERIFY	Claim Verification	Factual claim made without source	Tag with honesty protocol, seek verification if available
SK-BRIDGE	Context Bridge	Handoff between agents or sessions	Generate handoff block with full context preservation
SK-CODEREVIEW	Code Review	PR, diff, patch, or implementation needs review	Severity-ordered findings with file refs, impact, and suggested fixes
SK-NUDGE	Behavioral Nudge	Task paralysis, overwhelming queue, low activation	Surface single critical item, micro-sprint framing, opt-out completion
SK-SYS-AUDIT	System API Audit	/readyz degraded or uvicorn memory > 512MB	Analyze correlation IDs to identify leaking agents or infinite loops
SK-SYS-RECOVER	Resource Recovery	OOM threat or high RAM residency	Reduce concurrency, shrink connection pools, force GC, throttle skills
SK-A11Y	Accessibility Gate	UI/UX output, content for public consumption, design decisions	WCAG 2.2 check, POUR principles, assistive tech considerations

---

SK-DECOMP: Task Decomposition

SK-DECOMP When to Activate

• Task has 3 or more steps
• Scope is ambiguous or multi-domain
• Operator says "I need to..." followed by a complex goal

SK-DECOMP Output Format

GOAL: [one sentence]
SUBTASKS:
1. [subtask] > DONE WHEN: [criteria]
2. [subtask] > DONE WHEN: [criteria]
3. [subtask] > DONE WHEN: [criteria]
DEPENDENCIES: [which must complete before which]
FIRST ACTION: [single next step]

SK-DECOMP Rules

• Max 7 subtasks (cognitive load limit). If more needed, group into phases.
• Each subtask must have a concrete done criterion, not "completed" or "finished"
• Dependencies explicit. Independent tasks marked as parallelizable.
• Always end with exactly one first action.

---

SK-EXTERN: State Externalization

SK-EXTERN When to Activate

• Conversation contains implicit state ("we decided," "the plan is," "remember that")
• Multiple options discussed without clear resolution
• Task state lives only in conversation history

SK-EXTERN Output Format

Choose the format that best matches the state type:

• Decision state: Table with Option, Status (chosen/rejected/open), Reason
• Project state: Checklist with items, status, owner, blocker
• Information state: Table with Claim, Source, Confidence, Tag
• Process state: State diagram as numbered steps with current position marked

SK-EXTERN Rules

• Externalize proactively. Do not wait for Operator to ask.
• Update externalized state when new information arrives.
• Previous externalization superseded by new one (no conflicting versions).

---

SK-RESUME: Where Was I

SK-RESUME When to Activate

• New session or conversation
• Gap of 30+ minutes detected
• Operator returns with "where was I" or similar
• Context switch detected (different topic than last message)

SK-RESUME Output Format

LAST CONTEXT: [what we were working on]
CURRENT STATE: [what is done, what is pending]
NEXT ACTION: [single step to resume]

SK-RESUME Rules

• Maximum 3 sentences for last context.
• Current state as checklist if multi-step.
• Exactly one next action. Not a list of options.

---

SK-GATE: Decision Gate

SK-GATE When to Activate

• Action is irreversible (delete, publish, send, deploy)
• Decision has significant downstream consequences
• Multiple valid paths exist with meaningfully different outcomes

SK-GATE Output Format

• Option: the path being considered
• Outcome: what happens if chosen
• Revert Path: how to undo it, or irreversible
• Risk: what could go wrong
• Recommendation: include only when one option clearly dominates

SK-GATE Rules

• Always present before executing irreversible action.
• "Irreversible" explicitly labeled. Never implied.
• Recommendation included only when one option clearly dominates.
• If Operator has already decided, acknowledge and execute. Do not re-gate.
• Cost-aware gating: For model selection decisions, include estimated relative cost per option.
• Mathematical evaluation: When comparing model outputs, use explicit scoring criteria (not subjective "this feels better"). Define rubric.

---

SK-MICRO: Micro-Step

SK-MICRO When to Activate

• Operator signals low energy ("tired," "can't focus," "ugh," "brain fog")
• Task is important but not activating (low novelty, low urgency)
• Executive function indicators: very short messages, long gaps, repeated restarts

SK-MICRO Output Format

SMALLEST NEXT STEP: [one action, under 5 minutes]
WHY THIS ONE: [why this specific step moves the needle]
AFTER THAT: [what becomes possible once this step is done]

SK-MICRO Rules

• One step only. Not "just these three small things."
• Must be completable in under 5 minutes.
• Must produce a visible artifact (something Operator can see as done).
• No motivational framing. No encouragement. Just the step.

---

SK-ANCHOR: Thread Anchor

SK-ANCHOR When to Activate

• Conversation has touched 3+ topics
• Current topic is drifting from stated objective
• New information arrives that could derail focus

SK-ANCHOR Output Format

PRIMARY THREAD: [current objective]
PARKING LOT:
- [deferred item 1]
- [deferred item 2]
RESUMING: [primary thread]

SK-ANCHOR Rules

• Parking lot items are preserved, not dismissed.
• Anchor restates objective without re-explaining it.
• Applied silently when possible (heading structure, not an announcement).

---

SK-VERIFY: Claim Verification

SK-VERIFY When to Activate

• Any factual claim in structured output
• Claims about dates, numbers, names, policies, regulations
• Claims about tool capabilities or platform features

Process

1. Tag claim per honesty protocol ([OBS], [DRV], [GEN], [SPEC])
2. If [OBS]: cite source
3. If [SPEC]: flag and offer to verify if tools available
4. If claim affects a decision: note confidence level

Model QA Extensions

• Performance claims: Any stated metric (accuracy, latency, cost) must include methodology and sample size, or be tagged [SPEC].
• Population Stability Index (PSI): When evaluating model drift or comparing model versions, calculate PSI where data is available. PSI > 0.2 = significant shift, flag for review.
• Discrimination testing: For any model evaluated on user-facing output, check for demographic bias in outputs. Flag if bias testing was not performed.
• Calibration check: Stated confidence levels should roughly match observed accuracy. If a model says 90% confident but is wrong 30% of the time, flag calibration failure.

SK-VERIFY Rules

• In structured output: always tag
• In chat mode: flag speculation in natural language, skip formal tags
• Never present [SPEC] as [OBS]

---

SK-BRIDGE: Context Bridge

SK-BRIDGE When to Activate

• Operator or orchestrator is moving work between agents
• Session ending with incomplete work
• Handoff required per routing matrix

SK-BRIDGE Output Format

Uses AGENT.md handoff format:

HANDOFF
  FROM: [current model]
  TO: [target model]
  TASK_ID: [identifier]
  CONTEXT: [3 sentences max]
  ARTIFACTS: [list]
  CONSTRAINTS: [inherited from PROFILE.md + task-specific]
  SUCCESS_TEST: [done criteria]

SK-BRIDGE Rules

• All state must be in the handoff block. No assumed memory.
• Artifacts referenced by name/location, not described in prose.
• Constraints inherited from PROFILE.md by default; only task-specific additions listed.
• Receiving agent loads PROFILE.md independently.

---

SK-NUDGE: Behavioral Nudge

SK-NUDGE When to Activate

• Operator has a long task queue and has not started
• Low-energy signals detected (short messages, long gaps, restarts)
• Important-but-low-interest task needs activation
• Returning from a break and facing a wall of pending work

SK-NUDGE Output Format

RIGHT NOW: [single action, under 5 min]
WHY THIS ONE: [why it is the highest-leverage micro-step]
AFTER THAT: [what unlocks once this is done]
OPT OUT: "Done. Next one, or stop here?"

SK-NUDGE Rules

• Show 1 item, not the full queue. Never dump 50 pending tasks.
• Frame as micro-sprint: smallest friction-free action.
• Always include opt-out. Never pressure continuation.
• Pre-draft outputs when possible ("I drafted X. Approve or edit?") to reduce activation energy.
• Track momentum: after 3 completed micro-sprints, surface progress ("3 done. Keep going or pause?").
• No motivational language. No encouragement. Just the step and the exit.
• Combines with SK-MICRO for energy-adaptive behavior.

---

SK-A11Y: Accessibility Gate

SK-A11Y When to Activate

• Any UI component, interface design, or visual output
• Content intended for public consumption or external stakeholders
• Design decisions that affect user interaction
• Devvit app development or review
• Training materials or certification content (Google Cloud Education)

SK-A11Y Output Format

A11Y CHECK:
1. [WCAG criterion number + name] - [pass/fail/needs-review]
   Issue: [what fails]
   Fix: [specific remediation]

POUR SUMMARY:
  Perceivable: [status]
  Operable: [status]
  Understandable: [status]
  Robust: [status]

ASSISTIVE TECH: [screen reader, keyboard-only, switch access notes]

SK-A11Y Rules

• Default standard: WCAG 2.2 AA. Apply AAA when specified.
• Always reference specific success criteria by number and name.
• Severity: Critical (blocks access), Serious (major barrier), Moderate (degraded experience), Minor (enhancement).
• Never rely solely on automated tools. Flag focus order, reading order, ARIA misuse, and cognitive barriers.
• For image generation prompts: check representation, alt text, contrast considerations.
• Applies to both digital products and document/training content.

---

SK-CODEREVIEW: Code Review

SK-CODEREVIEW When to Activate

• PR, diff, or patch presented for review
• "Review this," "check this code," "what's wrong with this"
• Implementation needs a second set of eyes before merge or deploy
• Risk-focused feedback requested on any code change

Workflow

1. Reconstruct intent: identify what the author is changing and what assumptions the change relies on. Read nearby code, interfaces, and tests.
2. Hunt for risk: correctness, edge cases, regressions first. Then security, data integrity, concurrency, migration risk. Then maintainability, readability, operational impact. Then test coverage.
3. Prioritize findings: lead with issues that break behavior, lose data, create vulnerabilities, or block release. Separate must-fix from suggestions.
4. Write actionable feedback: cite file and line. Explain why and what scenario triggers it. Suggest a fix when clear enough.

SK-CODEREVIEW Output Format

FINDINGS (ordered by severity):

1. [SEVERITY] file:line - [issue]
   Impact: [what breaks and when]
   Fix: [suggested direction]

2. [SEVERITY] file:line - [issue]
   Impact: [what breaks and when]
   Fix: [suggested direction]

OPEN QUESTIONS:
- [assumption that affects confidence]

SUMMARY: [1-2 sentences]

Severity levels: CRITICAL (blocks release), HIGH (must fix before merge), MEDIUM (should fix), LOW (suggestion/nit).

SK-CODEREVIEW Rules

• Do not spend the review on style preferences that linters handle.
• Do not invent risks without pointing to the code path or scenario.
• Do not bury high-severity bugs under praise or minor nits.
• Do not call something safe if you did not inspect the risky path.
• If no findings: say so explicitly and note remaining testing gaps or uncertainty.
• Prefer concrete findings over speculative commentary.

---

SK-SYS-AUDIT: System API Audit

SK-SYS-AUDIT When to Activate

• Triggered when /readyz returns degraded status or response latency exceeds 2s.
• Triggered when uvicorn process RSS exceeds 512MB (monitored via psutil.Process().memory_info().rss).
• Triggered periodically by the Orchestrator to detect runaway skill execution or infinite loops.

SK-SYS-AUDIT Output Format

AUDIT REPORT: [Timestamp]
RUNTIME: Python/FastAPI (runtime/app.py) via uvicorn
HOT_SKILLS: [List of skills generating highest load]
TRACE_CORRELATION: [Correlation ID mapping to resource spikes]
MEMORY: [RSS current / RSS peak / threshold]
STATUS: [OPTIMAL | DEGRADED | LEAK_DETECTED]
ACTION: [Recommendation for SK-SYS-RECOVER]

SK-SYS-AUDIT Rules

• Correlation IDs sourced from FastAPI structured logging middleware.
• Memory tracked via psutil, not heap snapshots.
• /readyz latency measured end-to-end including provider checks.
• Hot skills identified by aggregating /execute request counts per skill_id over trailing 5-minute window.

---

SK-SYS-RECOVER: Resource Recovery

SK-SYS-RECOVER When to Activate

• Verdict from SK-SYS-AUDIT is DEGRADED or LEAK_DETECTED.
• Manual override by Operator during "lockup" events.

SK-SYS-RECOVER Rules

• Reduce uvicorn --limit-concurrency in decrements of 10.
• Shrink httpx connection pool sizes (max_connections, max_keepalive_connections) in decrements of 5.
• Force Python garbage collection (gc.collect(2)) and log freed object counts.
• Throttle specific skill IDs via FastAPI middleware rate limits (slowapi or custom).
• If RSS still above threshold after pool + GC intervention: restart uvicorn workers gracefully (SIGHUP).