Prepare Next Steps
Purpose
Pre-flight reflection run once per phase, before autonomous build. User decomposes plans into phases of 8-10 work items. Surfaces decisions, risks, and unknowns from a domain expert's view so the user can resolve them before triggering execution.
Trigger
Phrases: "prepare next steps", "prep next steps", "before we build", "pre-flight", or near variants. Also when the next user message is plausibly a build trigger and no pre-flight has been produced. Do not run mid-build.
Role Selection
Derive the most fitting expert role from the phase domain (e.g. Database Engineer for schema, UX Designer for interaction). Apply that role's domain priorities throughout sections 1-5. For multi-domain phases, dispatch panel mode.
Subagent Dispatch
Skip if subagents unavailable; fall back to single-thread.
Mode: single auditor (default) | panel of 2-3 parallel subagents with distinct domain roles (multi-domain or novel phases). Panel must include at least one structurally adversarial role (Security Engineer, SRE, future maintainer) when the phase has operational, security, or longevity implications.
Input: plan, work items, schemas/contracts only. No conversation, no prior decisions, no main-thread candidates.
Output: unranked candidates for decisions, concerns, questions, and dimensions.
Synthesis: main thread generates own candidates, merges (semantic dedup), compresses, ranks under selected role. Subagent-only items carry [fresh-context].
Response Header
Analysing Topic... ([keywords]), Selected best Topic Expert:
[Role]
Subagents dispatched: [single | panel: A, B, C | none]
Keywords: 3-6 short tags.
Output Format
Emit five sections in order, headers verbatim, counts fixed for 1-3:
- Top 4 Design Decisions (Ranked) - rank by irreversibility x blast radius x coupling x role. Per item: decision, 2-4 options, Claude's default if unanswered,
[fresh-context] marker if applicable.
- Top 4 Concerns about Current Design (Ranked) - rank by probability x impact x late-detection cost x role. At least one concern must be generated by pre-mortem inversion (assume the phase failed catastrophically post-deployment; identify the most likely cause). Per item: concern, failure mode, cheapest mitigation, marker if applicable.
- Top 6 Questions about the Work (Ranked) - rank by information value (work items unblocked per answer). Prefer scope, contracts, data shape, acceptance criteria. Per item: question, dependent items, best guess, marker if applicable.
- Other Information Needed for Autonomous Completion - free-form, no duplicates of 1-3. Typical: credentials, fixtures, schemas, scope boundaries, stop conditions, acceptance criteria, secondary domains. If empty: "None - proceeding is safe with current context."
- Unconsidered Solution Dimensions - fresh-context findings that reframe rather than refine: alternative architectures, wrong-unit-of-work, invalidating assumptions, cross-cutting concerns the plan treats as out of scope but which materially shape the approach, and constraint challenges (would relaxing any stated constraint eliminate the hard part?). If empty: "None - current plan dimensions hold under fresh-context audit." If subagents unavailable: "Skipped - subagent dispatch unavailable."
Reasoning & Ranking
- Materiality gate (applied before ranking): drop candidates that fail a regret/cost test. Decisions - would picking wrong cost more than resolving now? Concerns - would the failure cost more than mitigating now? Questions - would Claude's best guess be materially wrong often enough to warrant asking? If no, drop. When uncertain whether an item is material, drop. The cost of an unnoticed minor concern is lower than the cost of the user triaging a nitpick.
- Compress merged candidates before ranking; outliers do not displace stronger candidates.
- Project failures forward - propagating concerns rank higher than localised ones.
- Apply role-specific priorities, not generic ones.
- Surface sequence dependencies in section 1 or 4.
- Rankings are ordinal; if material candidates fall short of count, surface the gap rather than pad with non-material items.
- One issue per section; no cross-section duplication.
[fresh-context] items downranked only by domain priority, never familiarity.- Re-evaluate role and dispatch mode on domain shift; announce the change.
Format
- Numbered lists for sections 1-3.
- 1-3 sentences per item. No prose paragraphs.
- No emojis, no encouragement, no customer-service framing.
Handoff
After briefing, stop. Wait.
Build trigger: any message whose primary intent is to instruct execution start - "go", "proceed", "build it", "ship it", "execute", "do it", "start", or context variants. Match by intent.
- Clear trigger -> proceed.
- Answer without trigger (full or partial) -> wait; if partial, re-emit only unresolved items.
- Ambiguous -> ask once. Do not assume.
If section 5 reframings exist and the user triggers without addressing them, halt and ask: accept, reject, or defer? Reframings invalidate; do not carry untreated.
1---2name: prepare-next-steps3description: Trigger when user asks to "prepare next steps", "prep next steps", or signals readiness for autonomous build ("before we build", "pre-flight"). Produces a five-part briefing - 4 ranked design decisions, 4 ranked concerns, 6 ranked questions, an open information request, and unconsidered solution dimensions - from a domain-fitting expert role. When subagents are available, dispatches fresh-context auditors to surface candidates the main thread is anchored against. Apply when intent is to gate autonomous execution on prior reflection, even when phrasing is casual.4---56# Prepare Next Steps78## Purpose910Pre-flight reflection run once per phase, before autonomous build. User decomposes plans into phases of 8-10 work items. Surfaces decisions, risks, and unknowns from a domain expert's view so the user can resolve them before triggering execution.1112## Trigger1314Phrases: "prepare next steps", "prep next steps", "before we build", "pre-flight", or near variants. Also when the next user message is plausibly a build trigger and no pre-flight has been produced. Do not run mid-build.1516## Role Selection1718Derive the most fitting expert role from the phase domain (e.g. Database Engineer for schema, UX Designer for interaction). Apply that role's domain priorities throughout sections 1-5. For multi-domain phases, dispatch panel mode.1920## Subagent Dispatch2122Skip if subagents unavailable; fall back to single-thread.2324**Mode:** single auditor (default) | panel of 2-3 parallel subagents with distinct domain roles (multi-domain or novel phases). Panel must include at least one structurally adversarial role (Security Engineer, SRE, future maintainer) when the phase has operational, security, or longevity implications.2526**Input:** plan, work items, schemas/contracts only. No conversation, no prior decisions, no main-thread candidates.2728**Output:** unranked candidates for decisions, concerns, questions, and dimensions.2930**Synthesis:** main thread generates own candidates, merges (semantic dedup), compresses, ranks under selected role. Subagent-only items carry `[fresh-context]`.3132## Response Header3334> **Analysing Topic... ([keywords]), Selected best Topic Expert:**35> **[Role]**36> **Subagents dispatched:** [single | panel: A, B, C | none]3738Keywords: 3-6 short tags.3940## Output Format4142Emit five sections in order, headers verbatim, counts fixed for 1-3:43441. **Top 4 Design Decisions (Ranked)** - rank by irreversibility x blast radius x coupling x role. Per item: decision, 2-4 options, Claude's default if unanswered, `[fresh-context]` marker if applicable.452. **Top 4 Concerns about Current Design (Ranked)** - rank by probability x impact x late-detection cost x role. At least one concern must be generated by pre-mortem inversion (assume the phase failed catastrophically post-deployment; identify the most likely cause). Per item: concern, failure mode, cheapest mitigation, marker if applicable.463. **Top 6 Questions about the Work (Ranked)** - rank by information value (work items unblocked per answer). Prefer scope, contracts, data shape, acceptance criteria. Per item: question, dependent items, best guess, marker if applicable.474. **Other Information Needed for Autonomous Completion** - free-form, no duplicates of 1-3. Typical: credentials, fixtures, schemas, scope boundaries, stop conditions, acceptance criteria, secondary domains. If empty: "None - proceeding is safe with current context."485. **Unconsidered Solution Dimensions** - fresh-context findings that reframe rather than refine: alternative architectures, wrong-unit-of-work, invalidating assumptions, cross-cutting concerns the plan treats as out of scope but which materially shape the approach, and constraint challenges (would relaxing any stated constraint eliminate the hard part?). If empty: "None - current plan dimensions hold under fresh-context audit." If subagents unavailable: "Skipped - subagent dispatch unavailable."4950## Reasoning & Ranking5152- **Materiality gate (applied before ranking):** drop candidates that fail a regret/cost test. Decisions - would picking wrong cost more than resolving now? Concerns - would the failure cost more than mitigating now? Questions - would Claude's best guess be materially wrong often enough to warrant asking? If no, drop. When uncertain whether an item is material, drop. The cost of an unnoticed minor concern is lower than the cost of the user triaging a nitpick.53- Compress merged candidates before ranking; outliers do not displace stronger candidates.54- Project failures forward - propagating concerns rank higher than localised ones.55- Apply role-specific priorities, not generic ones.56- Surface sequence dependencies in section 1 or 4.57- Rankings are ordinal; if **material** candidates fall short of count, surface the gap rather than pad with non-material items.58- One issue per section; no cross-section duplication.59- `[fresh-context]` items downranked only by domain priority, never familiarity.60- Re-evaluate role and dispatch mode on domain shift; announce the change.6162## Format6364- Numbered lists for sections 1-3.65- 1-3 sentences per item. No prose paragraphs.66- No emojis, no encouragement, no customer-service framing.6768## Handoff6970After briefing, stop. Wait.7172Build trigger: any message whose primary intent is to instruct execution start - "go", "proceed", "build it", "ship it", "execute", "do it", "start", or context variants. Match by intent.7374- Clear trigger -> proceed.75- Answer without trigger (full or partial) -> wait; if partial, re-emit only unresolved items.76- Ambiguous -> ask once. Do not assume.7778If section 5 reframings exist and the user triggers without addressing them, halt and ask: accept, reject, or defer? Reframings invalidate; do not carry untreated.
Kalabint