Writing Specs

You are generating a Research Spec — a document that is simultaneously human-readable and machine-executable. Another CC session will later read this spec and execute it step by step.

Prerequisites

Before invoking this skill, the following MUST exist:

• A confirmed North Star statement
• A structured ResearchBrief
• Both preserved in a context-checkpoint file

Flow

Step 1: Read Research Catalog

Read skills/research-catalog/SKILL.md in its entirety. Internalize:

• All available campaigns and their strategies
• The dependency relationships between campaigns
• Pre-conditions for each campaign

Step 2: Structured Questioning

Invoke these 3 SOPs sequentially. Each asks 2-3 focused questions:

1. scope-clarification — research boundaries, depth vs breadth
2. campaign-selection — which campaigns to include/emphasize/skip
3. constraint-elicitation — time budget, existing knowledge, hard constraints

When the outline includes an experiment-execution stage, default to SUGGESTING ara-from-context as a closing stage (compile the results into an ARA); the user may decline. Without experiment-execution, still offer it if the user wants the research packaged. ara-from-context is the 10th optional package — never a forced tail.

Step 3: Pipeline Outline

Synthesize the North Star, ResearchBrief, and user answers into a 5-10 line outline:

Stage 1: [campaign] ([strategies]) — [topic/focus]
Stage 2: [campaign] ([strategies]) — [topic/focus]
...
Stage N: experiment-execution (experiment-design) — [topic]
Stage N+1: ara-from-context — compile the research into an ARA (OPTIONAL closing stage; include only if the user wants the results packaged for agent reproduction; user may remove)

Present this outline to the user. Wait for confirmation. User may adjust stages, reorder, add, or remove.

Step 4: Write Full Spec

Expand the confirmed outline into a complete Research Spec. Follow this schema exactly:

Spec Header

# Research Spec: <Topic>

> Generated: YYYY-MM-DD
> North Star: <one sentence>
> Scope: <N> stages, estimated <M> sessions
> Source: de-anthropocentric-research-engine

Global Sections

• Global Context Protocol (context-init/checkpoint rules)
• Global Execution Rules (±10% deviation, backtrack confirmation)
• Global Backtrack Conditions

Per-Stage Structure

For EACH stage, write ALL of these fields:

• Objective: What this stage accomplishes
• Expected Input: What context is available from prior stages
• Focus Areas: Specific aspects to emphasize
• Recommended Combination: campaign → strategy-A, strategy-B
• Completion Criteria: Quantified threshold (numeric or objectively verifiable)
• Backtrack Condition: if condition], → Stage N (requires user confirmation)
• Execution Steps: Checkbox items (context-init, each strategy, checkpoints)

Granularity Rules

• Name specific strategies (not tactics/SOPs — those are CC's choice at execution time)
• Describe what topic/aspect each strategy addresses
• Specify what prior context is available and how to use it
• Quantify ALL completion criteria (no vague "sufficient" or "adequate")
• Focus Areas tell CC what to prioritize within the campaign's scope

ARA Closing Stage (when present)

When the outline includes an ara-from-context stage, write its Execution Steps so the stage opens with an explicit user-decision gate, NOT an automatic transition:

• First step: after experiment-execution's evaluation reports PASS, present to the user:

"实验评估通过,是否进入 ARA 成文?(compile results into an ARA)" — proceed only on approval.

• Do NOT hard-code which EE evaluation skill produces that PASS; the executing agent reads

the spec's experiment-execution stage and the research-catalog to choose it at run time.

• The gate is an ordinary natural-language Execution Step (spec = trace). Do not invent a

new spec field, and do not reuse the Backtrack Condition mechanism (that is for retreat, this is a forward go/no-go decision).

• Remaining steps: invoke ara-from-context (context-review → compile-and-review), producing

ara/ + ara/level2_report.json. Note the external prerequisite: npx @ara-commons/ara-skills must be installed (compiler + rigor-reviewer).

Step 5: Spec Self-Review

Invoke spec-self-review SOP. This is MANDATORY and cannot be skipped.

Step 6: User Review

Present the completed spec to the user for review. Wait for approval or change requests. If changes requested, revise and re-run self-review.

Output

Save the spec to: docs/de-anthropocentric/specs/YYYY-MM-DD-<topic>-spec.md

Inform the user: "Spec complete. To execute, invoke executing-specs with the path to this spec file."

<!-- BEGIN available-tables (generated) -->

Available SOPs

Optional, no fixed order; the final leaf is always a sop.

| SOP | When to use | | --- | --- | | campaign-selection | Structured questioning SOP to determine which campaigns to include, emphasize, or skip. Used during spec generation. | | constraint-elicitation | Structured questioning SOP to identify practical constraints that shape the research spec. Used during spec generation. | | research-catalog | Capability menu for the research engine. Lists the 10 freely-composable research packages, what each does, when to reach for it, and a pointer to its full skill table. Read this after north-star crystallization to decide which packages to use — no fixed order. Also serves as the skill-index (capability map). | | scope-clarification | Structured questioning SOP to determine research boundaries, depth, and breadth. Used during spec generation. | | spec-self-review | Quality gate for Research Specs. Checks for placeholders, consistency, scope, ambiguity, context protocol, and quantification. Mandatory before user review. |

<!-- END available-tables (generated) -->