align

/align — share understanding before acting

The most expensive failures aren't bad outputs — they're outputs built on a wrong premise. By the time the user sees prose, you've silently committed to: your reading of the request, your chosen approach, your interpretations of source, your register and format choices, and a hundred small fill-ins for things that weren't specified. Each commitment looks authoritative once written.

This skill inverts that. Before producing any artifact, produce a manifest — a structured, scannable summary of every silent commitment. The user reacts. Drafting happens after.

The manifest

Output exactly five blocks. Target <60 seconds to read. If a block balloons, the source is thin or the scope is too broad — surface that, don't pad.

Tag every item:

• [grounded] — directly stated in the prompt or source.
• [inferred from "X"] — derived from grounded content; show the chain.
• [mine] — autonomously introduced; no derivation from source.

Tags answer different questions: grounded → confirm, inferred → correct the premise, mine → accept or override.

Block 1 — My read of the task

Paraphrase the request. State the audience. State why this matters / what it's for. Note any timing constraints. Form details (length, voice, structure) belong in Block 4.

Block 2 — My approach

Describe the method. Include scope: what to do AND what to explicitly NOT do. Surfaces silent strategy commitments — asked for a draft, got a strategy doc; asked for a bug fix, got a refactor.

Block 3 — What I'm working with

Two parts:

1. Source coverage — what's in the source, tagged.
2. What I'd invent if not stopped — explicit list of plausible-sounding details you'd otherwise fabricate.

The "what I'd invent" callout is the anti-fabrication bell — forcing enumeration makes invention visible; details that would have slipped uncited get surfaced as choices.

Format:

- Source fact `[grounded]`
- Derivation from source `[inferred from "..."]`

**What I'd invent if not stopped:**
- A plausible detail not in source
- A specific number or date the source doesn't mention

Block 4 — How I'd render it

State the shape (length, structure, container) and a sample of the actual output. The sample carries register, voice, conventions — don't list style attributes the sample reveals on its own. Catches silent voice drift.

Format:

- Shape: [length, structure, container]
- Sample: *"[an actual opening line of the output, in the intended voice]"*

Block 5 — Blocking questions

Only forks where any default would be unsafe. Different from gaps you'd fill with defaults — those go in Block 3 as "what I'd invent." If there are none, say so. Don't pad.

After the manifest

Stop. Wait for the user to confirm, correct, or re-scope. Once confirmed, re-read the manifest before drafting — committed items are load-bearing. If something needs to change during drafting, pause and surface it: "I'd like to add X — wasn't in the manifest. OK?" Don't drift silently.

Common pitfalls

• Manifest theater. Clean manifest, then drift in the draft. Re-read before drafting; surface deviations.
• Padding to look thorough. Inflating blocks to fill space. If approach is trivial, one bullet is enough.
• Hiding inventions in [inferred]. If there's no derivation chain to point at, it's [mine].
• Skipping the "what I'd invent" callout. The anti-fabrication bell. Empty means either nothing's being filled in, or the looking has stopped.
• Treating the manifest as a contract. It's a checkpoint. Drift is allowed when surfaced, not when silently absorbed.
• Confusing align with clarify. Align commits to specific reads; clarify asks where you're stuck. Different modes.
• Summarizing from memory instead of reading source. Block 3 requires actually reading what's there. If source is too long, say so explicitly.