Agent Skills: Amplify

Amplify

Turn a rough prompt into a high-signal one. Detect the domain, inject the standards that domain demands, score it against a rubric — then offer to run it.

Dispatcher and reviewer — Opus 4.8 (thinking-tier). Searcher/Writer — Sonnet 4.6 (worker-tier).

Amplify does one thing well: it rewrites the prompt you give it into the single strongest version, then hands it off. It does not write code itself — it produces a prompt other skills (or you) execute.

Per-Step Agent Map (DOCTRINE rule 12)

Every substantive step dispatches at least one Agent. Atomic steps (DOCTRINE 12.2.8) are a single Worker → Reviewer pair with no independent angles to fan out.

| Step | Status | Worker tier | Thinking tier | Notes | |---|---|---|---|---| | 1 — Read intent | Atomic (12.2.8) | Searcher (Sonnet) — domain signals + project rules | Analyst (Opus) — triage + gap analysis + persona selection | Single Worker → Reviewer; reads CLAUDE.md, AGENTS.md, .hyperflow/memory/* | | 2 — Amplify | Atomic (12.2.8) | Writer (Sonnet) — draft the enhanced prompt | Reviewer (Opus) — rubric score, one targeted revision | Writer drafts → Opus scores against the 8-dim rubric → revise once if any dim < 4 | | 3 — Present + handoff | §12.1 inline | — | — | Print the amplified prompt + rationale; fire the handoff gate (AskUserQuestion) |

Inputs

• $ARGUMENTS — the raw prompt to amplify. If empty, amplify the user's previous message in the conversation.
• No flags. Amplify always produces the single best version (per the design decision); the rubric, not a flag, governs depth.

Flow

Step 1 — Read intent (atomic · 12.2.8)

Atomic — a single Searcher → Analyst pair. No parallel angles: understanding one prompt is one scope.

1. Dispatch Searcher — detecting prompt domain + gathering project rules (Sonnet). It returns:

   • Domain signals — does the prompt concern frontend/ui/creative, api/db backend, mobile, security, performance, refactor/bugfix/test, devops, docs? (A prompt can span several.)
   • Project rules — read, when present: CLAUDE.md, AGENTS.md, .hyperflow/memory/conventions.md, .hyperflow/memory/project-decisions.md, .hyperflow/memory/anti-patterns.md. Extract any rule that should constrain the prompt.
   • The raw prompt's gaps — what a senior engineer would have to guess to act on it.

2. Dispatch **Analyst** — triaging intent + selecting personas (Opus). It returns { domain[], intent, ambiguities[], personas[], project_rules[] } — the matching hyperflow persona(s) from ../hyperflow/personas-A.md / personas-B.md, plus the project rules to layer on top.

If the prompt is too ambiguous to amplify without a decision the rewrite can't make (ambiguity high on a dimension that changes the deliverable), the Analyst flags it — Step 3's handoff gate surfaces it as a clarifying note rather than guessing.

Step 2 — Amplify (atomic · 12.2.8)

Atomic — a single Writer → Reviewer pair with a one-shot revision loop.

1. Dispatch Writer — drafting the amplified prompt (Sonnet) with the Step 1 output. The Writer rewrites the raw prompt into the single strongest version following the skeleton in references/prompt-rubric.md: role framing · precise task · context · constraints (persona doctrine + project rules) · output spec · out-of-scope. Economy is a constraint — enhance to the level the task warrants, never inflate a one-line ask into a spec.

2. Dispatch **Reviewer** — scoring against the prompt-quality rubric (Opus) with the draft. It scores all 8 dimensions (see references/prompt-rubric.md) 1–5. Verdict:

   • All dimensions ≥ 4 → PASS. Ship the draft.
   • Any dimension < 4 → NEEDS_REVISION with the specific dimensions + what's missing. The Writer revises once with the findings injected, then the result ships regardless (no infinite loop — DOCTRINE rule 14 NEEDS_REVISION cadence: surface after the second pass).

The Reviewer also produces the rationale — a 2–4 line "what changed and why," naming the domain doctrine and project rules it injected.

Step 3 — Present + handoff (inline · §12.1)

Trivial-inline — no Agent dispatch. The orchestrator prints, then gates.

1. Print the amplified prompt in a single copy-ready fenced block.

2. Print the rationale beneath it — what changed, which persona standards + project rules were injected, and any ambiguity the Analyst flagged.

3. Fire the handoff gate — AskUserQuestion:

   Run this amplified prompt now?

   • Send to spec (recommended) — start the chain design-first; spec brainstorms + designs with the amplified prompt, then auto-chains to scope → dispatch
   • Send to scope — the approach is clear, go straight to decomposition (/hyperflow:scope)
   • Send to dispatch — a task file already exists (/hyperflow:dispatch)
   • Copy only — keep the prompt, run nothing

   This is a multi-option gate (4 options) → it carries the (Recommended) marker on Send to spec (DOCTRINE rule 8 — named-workflow choice).

   On a Send to … selection, invoke the chosen skill via the Skill tool, passing the amplified prompt as the argument. Spec is the default because a freshly-amplified prompt is a design starting point — spec's brainstorming + section-by-section design is exactly what it feeds. On Copy only, stop — the prompt is already printed for the user to take.

   Codex fallback: if the host does not expose the AskUserQuestion popup UI, print the same gate as a Hyperflow Question chat block with the four numbered choices and wait for the user's reply. Do not auto-select Send to spec.

Iron Rules

• Amplify never writes code. It produces a prompt; downstream skills execute it. If the user wanted code, the handoff gate routes there.
• Project rules win on conflict. A rule in CLAUDE.md / AGENTS.md / .hyperflow/memory/ overrides the generic persona standard — it is the user's explicit instruction.
• Economy is mandatory (rubric dim 8). Never inflate a trivial prompt. The rubric enhances to the task's level, not beyond.
• Failure recovery (rule 14). Worker/Reviewer errors and the NEEDS_REVISION cadence follow ../hyperflow/failure-recovery.md.
• No AI attribution in the amplified prompt or rationale — describe the work, never the author.

Doctrine

Shared rules in ../hyperflow/DOCTRINE.md. Personas in ../hyperflow/personas-A.md + personas-B.md. Rubric in references/prompt-rubric.md.