Skip to content

platform-capabilities.md

Latest commit

 

History

History
95 lines (58 loc) · 5.16 KB

platform-capabilities.md

File metadata and controls

95 lines (58 loc) · 5.16 KB

Platform capabilities — canonical reference for JAE

README and SKILL.md summarize this document; they must not contradict it.

Execution modes (v2.1.x): EXECUTION_MODES.md — agent_supported (default) vs cowork_autonomous (CoWork, opt-in).
Scope law: MANDATORY_EXCLUSIONS.md.

Claude.ai (web chat; Skills or Project Knowledge)

In scope for JAE

  • Full eight-phase workflow sequentially in one conversation.
  • Scoring gate after every phase; explicit confirmation before Phase 4 (first application text).
  • Tools/MCP available in that chat session (per host policy): web_search, web_fetch, and others probed in A0.
  • Agent Skills (where the product offers them): upload skill ZIP per product UI — see Anthropic documentation for current paths and plan requirements.
  • Execution mode: agent_supported only (no cowork_autonomous in this product surface).

Mandatory limitations

  • Phases do not run as parallel subagents; high-volume parallel discovery is slower than CoWork for the same authored parallelism.
  • “Agentic” here means in-conversation tool and MCP use within product limits — not the same as CoWork workspace-style execution.

Install (primary vs fallback)

  1. Primary: Skills upload (product-dependent naming — e.g. Customize → Skills → Create Skill).
  2. Fallback: Project Knowledge — upload SKILL.md, rules.json, automation-registry.json, and all files under references/ (see repo layout).

Claude CoWork

In scope for JAE (explicitly harnessed)

  • Parallel Tier-1 read-only work: Phases 0, 1, and salary research inside Phase 3, as parallel subagents, per SKILL.md and rules.json.
  • Sequential Phases 5 and 6; Phase 2 verdict before Phase 4 begins.
  • Tier 2 and Tier 3 automations: always on the main coordination thread after consent — never on a subagent.
  • Artifacts: checklist as static file updated per gate (references/skill-instructions/checklist-templates.md). Package output described as downloadable when the host supports file presentation.
  • MCP: BrowserBase, Playwright, email, cloud storage, calendar — only when connected and listed in A0 Capability Map.
  • Execution modes: Default agent_supported. User may opt into cowork_autonomous for larger host-driven chunks only with explicit opt-in and mandatory re-anchor to the same phase law after each chunk — see EXECUTION_MODES.md and A14–A16 in automation-registry.json.

Mandatory wording

  • Parallelism is authored in this skill and allowed by CoWork; the product does not “auto-route phases” without the skill’s rules.
  • Availability and plans of CoWork, connectors, and controls are product-owned; see Claude CoWork (Anthropic) and Cowork (claude.com). JAE does not hardcode a plan matrix here.

Scope (not “exclusion of CoWork”)

  • See MANDATORY_EXCLUSIONS.md: hard-prohibited JAE mis-claims, host-governed features under Mode 2 + re-anchor, registry-native automations for anything presented as a named Axx.

Manus

In scope for JAE (primary install — current product)

  1. Skills: Skills tab → + Add → Upload a skill (.zip / .skill) or Import from GitHub with this repository’s public URL.
  2. Fallback: Extract ZIP and upload workspace files; or paste rules.json as session instruction source.

Mandatory limitation

  • present_files may be INACTIVE on some Manus sessions. When INACTIVE, follow SKILL.md A0 — Mandatory artifact fallback (inline deliverable text + user save instructions). Never assume a file download primitive exists.
  • Execution mode: agent_supported only unless Manus later exposes an equivalent; default remains guided governance.

Hybrid patterns (both mandatory; one default)

Default: hybrid_chat_review

  • Run Phases 0 and 1 in CoWork for parallel discovery and company intelligence speed.
  • Copy the Company Intelligence Brief into Claude.ai.
  • Run Phases 2–7 in Claude.ai for sequential scoring gates and in-chat review of drafting.

When to recommend: User wants parallel discovery plus single-thread chat review for writing phases.

Secondary: cowork_end_to_end

  • Run Phases 0–7 entirely in CoWork when local artifacts, downloadable packages, or heavy file/browser work should stay in one workspace.

When to recommend: User prioritizes workspace files and CoWork-local delivery over chat-only review.

Optional: In CoWork, combine cowork_end_to_end with cowork_autonomous for execution chunks only under EXECUTION_MODES.md.

Machine-readable keys: rules.json → platform_notes.hybrid_chat_review, platform_notes.cowork_end_to_end.

Links