NextSteps

Append context-aware next-step suggestions after agent responses. Generates actionable follow-ups, surfaces unfinished tasks from memory, and includes creative lateral suggestions. Self-learns user preferences for count, categories, and format. Activate by: (1) any response when nextsteps is enabled, (2) user says /nextsteps, "next steps", "what should I do", "suggestions", "what now", (3) user wants to customize settings, (4) user says "disable next steps" or "enable next steps".

How It Works

This skill generates helpful next-step suggestions after responses. It reads user preferences from .nextsteps/PREFERENCES.md — if enabled: false, nothing is shown. When enabled, it produces exactly display-count suggestions following the pipeline below.

Quick Start

1. Check .nextsteps/PREFERENCES.md for user config — if enabled: false, stop (show nothing)
2. If .nextsteps/ does not exist, run the cold-start protocol (see COLD-START.md)
3. Generate next steps following the pipeline below
4. Append to your response in the correct format

Generation Pipeline

Follow these six steps to generate next steps:

Step 1 — Read Configuration

Read .nextsteps/PREFERENCES.md. Extract: enabled, display-count, preferred-categories, excluded-categories, format, show-footer. Category names in PREFERENCES.md use kebab-case (e.g., direct-follow-up). If the file is missing or unreadable, use defaults: enabled=true, display-count=5, format=standard.

Step 2 — Analyze Context

Determine: What did the user just accomplish or ask? What is the active topic? What is the session scope — quick fix, feature work, exploration, debugging, architecture? What is the user's likely next thought based on conversation trajectory?

Step 3 — Check Memory

Read .nextsteps/BACKLOG.md for unfinished items relevant to current context. Read PREFERENCES.md for topic affinities (STRONG/MODERATE/WEAK) and anti-preferences (topics and types to avoid). If memory files are unavailable, skip to Step 4 — memory is a bonus, not a gate.

Step 4 — Generate Candidates

Generate display-count suggestions using these six categories:

Icon	Category	Tier	Slot Rule
⚡	Direct Follow-up	STRONG	1 guaranteed
🔧	Actionable Task	STRONG	1 guaranteed
🔍	Deep Dive	MODERATE	1 when count ≥ 3
📋	Memory Recall	MODERATE	1 when relevant backlog exists
💡	Lateral / Out-of-the-Box	MODERATE	1 when count ≥ 3
✅	Quick Win	MODERATE	Fills remaining slots

STRONG categories get guaranteed slots. Remaining slots filled by MODERATE categories in round-robin. Respect excluded-categories. Prioritize preferred-categories. See CATEGORIES.md for detailed taxonomy and examples.

Step 5 — Self-Review Gate (CRITICAL)

Review every candidate against these core rules. Remove violators and regenerate:

1. No restating the obvious — don't echo what was just explained
2. No generic filler — ban "tell me more", "pros and cons?", "anything else?"
3. No hallucinated context — only reference things the user actually mentioned
4. No scope mismatches — match suggestion effort to session scope

Every suggestion should be: specific, actionable, non-obvious, contextually grounded, scope-appropriate, and differently framed from siblings. Full 11-rule checklist and violation examples in ANTI-PATTERNS.md.

Step 6 — Format and Present

Show exactly display-count items. Never more, never fewer.

Standard format (default for rich-text channels):

text

## ⚡ Next Steps

1. 🔧 **[Bold title]** — [Brief context explaining relevance]
2. 🔍 **[Bold title]** — [Brief context]
3. 📋 **Resume: [task from backlog]** — [When it was started]
4. 💡 **Consider: [creative lateral idea]** — [Why it matters]
5. ✅ **Quick win: [small action]** — [Time estimate]

_Your selections help me learn what matters to you._

Compact format (for TUI, character-limited channels):

text

⚡ Next: [1] Title | [2] Title | [3] Title

Token-budget rule: If your response is approaching the output token limit, switch to compact format with min-count items (default: 1). Reserve ~100 tokens for next steps when planning long responses. If even compact won't fit, place one inline suggestion before your final paragraph: (Next: [suggestion]).

Customization Detection

Before generating next steps, check if the user's message is a customization request. If it matches any of these patterns, process the config change and confirm:

• "show me N next steps" / "only N suggestions" → set display-count: N
• "disable next steps" / "stop showing suggestions" → set enabled: false
• "enable next steps" / "turn suggestions back on" → set enabled: true
• "compact format" / "shorter suggestions" → set format: compact
• "don't show backlog" → set include-backlog: false
• "hide the footer" / "no footer" → set show-footer: false
• "reset next steps settings" → reset all config to defaults
• "show next steps settings" → display current config

Update PREFERENCES.md immediately and confirm: "Got it — [description of change]." Log as [CONFIG-CHANGE] in HISTORY.md. See CUSTOMIZATION.md for full protocol.

Selection Tracking

After presenting next steps, detect what the user does on their NEXT message:

• User references a suggestion by number or content → log [SELECTED] #N category in HISTORY.md. Promote that category tier if MODERATE→consider STRONG.
• User asks something unrelated to any suggestion → log [IGNORED] all in HISTORY.md
• User gives negative feedback ("too many", "not helpful", "stop suggesting X") → log [FEEDBACK] and adjust per CUSTOMIZATION.md

Self-Improvement

Every 10th activation, run one learning experiment. Every 20 HISTORY.md entries, run a self-diagnostic. See SELF-IMPROVE.md for the full observe/hypothesize/experiment/validate cycle.

Key self-learning behaviors:

• If user only selects from top 2 items over 10 interactions → hypothesize lower count preference → experiment → validate
• If user ignores all suggestions 5+ times → trigger diagnostic, reset category weights
• After 5 validated experiments on count → set confidence HIGH, stop experimenting

Channel Adaptation

Detect the channel and adapt format:

• OpenClaw: Read channel from conversation metadata → adapt format per channel type
• VS Code / Rich text: Use standard format with icons and bold
• Terminal / TUI: Use compact format, no icons
• WhatsApp / Signal / iMessage: Use shortest compact form

All channels share the same .nextsteps/ state. Preferences learned on one channel apply everywhere.

Security Rules (CodeGuard)

These rules are always active. Derived from cisco/software-security (Project CodeGuard):

1. Never store secrets: .nextsteps/ files must never contain API keys, passwords, tokens, or credentials. If conversation context includes secrets, sanitize them from suggestions.
2. Validate before writing: Before writing any .nextsteps/ file, verify the path is within .nextsteps/ scope (no ../ traversal) and content contains no secret patterns (sk-, api_key=, password=, token=, secret=).
3. Minimize stored data: HISTORY.md stores titles and selection status only. PREFERENCES.md stores tiers and config only. BACKLOG.md stores brief descriptions only. Never store raw conversation text.
4. Enforce file limits: PREFERENCES.md ≤ 120 lines, HISTORY.md ≤ 50 entries, BACKLOG.md ≤ 30 items. Summarize overflow, don't truncate blindly.
5. Suggest .gitignore: On first activation, if .nextsteps/ is not in .gitignore, include it as a next step.

See SECURITY.md for the complete security protocol.

Error Recovery

• .nextsteps/ missing → generate next steps from conversation context alone; recreate files on next write
• PREFERENCES.md corrupted → recreate from defaults; preserve any readable sections
• HISTORY.md overflow → summarize oldest 25 entries into PREFERENCES.md tier adjustments, then clear them
• Any file read fails → proceed without that file; never block next-steps generation

Reliability Self-Check

If the response does not end with next steps and enabled is not false, append them using conversation context alone.