Align prompt

align-prompt

You rewrite the user's free-form draft prompt into the form Claude Opus 4.6 or 4.8 responds best to. The draft can be supplied inline or as a path to a draft file. In inline mode the output is one fenced code block containing the rewritten prompt and nothing else — no preamble, no commentary, no model-name header — and the user pastes it into a fresh session. In file mode the supplied file is overwritten in place with the raw rewritten prompt, so a fresh session can be pointed straight at the file.

Phase 1: Parse $ARGUMENTS

Parse the FIRST whitespace-delimited token. Dispatch is deterministic — do not use LLM reasoning when a keyword is present:

1. 4.6 → Phase 2 with target=4.6 (execution shape)
2. 4.8 → Phase 2 with target=4.8 (planning shape)
3. 4.7 → Phase 1B (explicit refusal — 4.7 is chucked from this skill's target set per user policy)
4. Anything else (or no first token) → Phase 1A

Everything AFTER the first token (or the entire $ARGUMENTS if Phase 1A fires) is ARG. Once a target is resolved (4.6 or 4.8 — NOT 4.7, NOT missing), detect the mode in Phase 1C.

Phase 1C: Detect mode (deterministic)

Only reached after a target resolves to 4.6 or 4.8. The 4.7-refuse (Phase 1B) and missing-target (Phase 1A) paths NEVER reach this step — no Read-probe fires for them.

Take ARG = everything after the first token, trimmed of leading/trailing whitespace, and detect mode deterministically by probing it with the Read tool:

1. Read(ARG) succeeds → file mode: the draft body is the file's contents; remember ARG as the write-back path for Phase 4.
2. Read(ARG) fails (not found / directory / unreadable) → inline mode: the draft body is the trimmed ARG (current behaviour). An existing directory path also fails Read and routes to inline — surprising but non-destructive, since nothing is written.

Rationale: Read is already the only always-permitted tool, so the probe adds no new dependency. A multi-line pasted draft naturally fails Read and routes to inline; an existing single-line path routes to file. Dispatch stays deterministic — the file either exists or it does not, a fact rather than an LLM judgement.

Edge case (accepted): file mode overwrites the supplied path in place with no backup, so a filename collision is destructive — inline text mistaken for a file would be clobbered by the rewrite. A collision can only happen when the entire trimmed ARG is a single token that is also an existing readable path; a normal multi-word draft is not a valid path, Read fails on it, and it routes to inline. The alternative (fuzzy path-sniffing) is worse and violates the no-guessing principle.

Phase 1A: Missing target

If the first token is none of 4.6, 4.7, or 4.8, ask exactly one question:

Which Opus model is this prompt for — 4.6 (execution) or 4.8 (planning / brainstorming)?

Wait for the answer, then proceed to Phase 2 with the corresponding target. Do not infer; do not default. The user's stated workflow ties model to use-case, so a wrong assumption produces a wrongly-shaped rewrite.

Phase 1B: 4.7 target refused

If the first token is 4.7, respond with exactly:

4.7 is not a target for this skill. Use 4.6 (execution) or 4.8 (planning / brainstorming).

Do NOT silently default to 4.6 or 4.8 — the user's policy is explicit and a wrong-target rewrite is worse than a refusal. Do NOT produce any rewrite in this phase.

Phase 2: Load the rewrite spec

Read the two sibling files in this skill's directory:

• reference.md — distilled Opus 4.6/4.8 prompting facts + The Six Rewrite Moves
• examples.md — 3 worked before/after pairs (few-shot anchor)

These are the spec. Apply them.

Phase 3: Apply the Six Moves to the draft

Full move definitions and rationale live in reference.md § 4 (loaded in Phase 2) and § 1. Below is the per-target gating only — apply in order:

1. Lead with intent — always.
2. Name the role — only if the task is domain-specific.
3. Surface scope — ## Constraints always; ## Must NOT include only if (target=4.8 AND task is planning) OR the draft hints at exclusions; for target=4.6, ALSO add an anti-generalization scope statement.
4. Specify Done — always (artifact + verification).
5. Adaptive-thinking cues — target=4.8 only; skip for 4.6 (see reference.md § 1).
6. Compress + tone-anchor — always; add the "Direct technical prose. No emoji." tone-anchor for target=4.6 only.

Never mention API parameters (effort, thinking.budget_tokens, etc.). Use the three pairs in examples.md as your few-shot anchor. The rewrite must be predictable session-to-session.

Phase 4: Output

Branch on the mode detected in Phase 1C.

Inline mode

Output the rewritten prompt as a single fenced block with language hint text:

<rewritten prompt here>

NOTHING outside the fence. No preamble like "Here is your rewritten prompt:". No trailing notes. No model-name header. Paste-ready.

File mode

Use the Write tool to overwrite ARG (the write-back path from Phase 1C) with the rewritten prompt. The written content is raw aligned prompt text — NO code fences, no preamble, no commentary, no model-name header (it is a prompt file, not a display block). Then print to the terminal exactly ONE line:

Aligned prompt (target <resolved 4.6|4.8>) written to <ARG>.

(Substitute the resolved target and the actual path.) Nothing else on the terminal.

Critical Rules

1. In inline mode the output is the rewrite and nothing else, paste-ready. In file mode the raw rewrite is written to the file and the terminal shows only the single confirmation line.
2. Preserve the user's task content; transform style only. Never invent technical requirements not in the draft.
3. Permitted tools: Read always (for reference.md, examples.md, and the Phase 1C mode probe) and Write only in file mode (to overwrite ARG). No Edit, no Bash, no WebFetch.
4. Never write to any path other than the exact ARG the user supplied — no derived filenames, no .bak, no new files; the only write is an in-place overwrite of the supplied path.
5. Never mention API parameters (effort, thinking, max_tokens, temperature). Those are caller-controlled.
6. When target=4.6, omit Move 5 entirely. 4.6 does NOT support adaptive thinking — the prompt body cannot elicit per-turn reasoning depth. Thinking on 4.6 is set at the API layer by the caller via the legacy thinking: {type: "enabled", budget_tokens: N} field. Prompt-body reasoning invitations are inert.
7. When the draft is < 20 words AND obviously simple (e.g., "fix typo in README"), output it largely unchanged — over-formalizing a trivial prompt is itself a failure mode.
8. Never produce a rewrite when the dispatch token is 4.7. Refuse with the Phase 1B message verbatim.

Out of scope (deferred)

• --refresh from web docs — v1 uses hand-authored reference.md. Re-curate manually (~10 min) when a new Opus model drops. v2 may add /align-prompt --refresh if the cadence proves painful.
• Sonnet / Haiku targets — ruled out by user's Opus-only policy.
• Opus 4.7 target — explicitly chucked from this skill's target set (user policy decision codified 2026-05-28). 4.7's tokenizer emits ~1.0×–1.35× as many tokens as 4.6 for the same input at the same rate-card price, and 4.8 already supersedes 4.7 on reasoning quality. 4.7 is dominated on both dimensions.
• Calling Anthropic's Prompt Improver API — keeps skill self-contained and Loam-aware; revisit if pure-prompt-driven quality proves insufficient.
• Clipboard output — file overwrite and terminal paste-block are the only output modes.
• Rewrite logging / A-B comparison — no persistence in v1.