Decision Forensics Skill

Most decisions are never stated — they precipitate out of a thread and everyone leaves with a different memory of them. This skill is the forensic pass: what was actually decided (if anything), who committed to what, and which disagreements got papered over rather than resolved. (To write a fresh decision going forward, use architecture-decision-record; this skill reconstructs one from the wreckage.)

What This Skill Produces

• The decision, stated cleanly — one sentence, even though nobody ever said it in one sentence — or the finding "no decision was reached"
• The commitments table — who agreed to do what, with the quote that binds them
• Reconstructed assumptions — what everyone silently took as true, labeled as reconstructed
• Dismissed options and papered-over questions — what was set aside, and what was never actually resolved
• A confidence note — how solid this reconstruction is and where it's guessing

Required Inputs

Ask for these if not provided:

• The thread — Slack export, email chain, or meeting notes/transcript (paste; timestamps and names preserved if possible)
• The cast (optional) — who's who: roles and decision authority; else infer from context and label
• What prompted the forensics (optional) — a dispute, an audit, onboarding someone — shapes emphasis, never conclusions

Framework: The Forensic Protocol

1. Anchor — find the moments where direction changed hands: an approval word ("ship it", "fine, let's"), an unobjected proposal followed by action talk, or authority going silent after a suggestion (silence + subsequent action = the most common decision form).
2. Attribute — every claim about a person gets a verbatim quote or close paraphrase with the original wording nearby. No quote, no attribution.
3. Excavate assumptions — what must everyone have believed for the exchange to make sense? Label each [reconstructed] — these were never said.
4. Separate resolved from papered-over — a question answered vs a question abandoned when someone changed the subject. The second list is usually the valuable one.
5. Grade the reconstruction — HIGH (explicit approval by named authority) · MEDIUM (unobjected proposal + consistent action) · LOW (inferred from fragments). State it.

Output Format

Decision Record (reconstructed): [one-sentence decision]

Source: [thread, date range] · Reconstruction confidence: HIGH/MEDIUM/LOW — [why]

The Decision

[One sentence. Or: "No decision was reached. The thread ends with X unresolved; subsequent action, if any, happened without recorded agreement."]

Commitments

Assumptions (all [reconstructed])

• …

Options Dismissed

Papered Over

[Questions raised and abandoned — each with who raised it and where the thread swerved.]

Quality Checks

• Every attribution carries a quote or near-paraphrase with original wording nearby
• Assumptions are labeled [reconstructed] — never presented as things people said
• The confidence grade matches the evidence type (explicit / unobjected / inferred)
• Papered-over questions are listed separately from resolved ones
• If no decision was reached, the record says so plainly

Anti-Patterns

• Do not attribute positions beyond what the text supports — quote it or drop it
• Do not resolve ambiguity by inventing agreement — surfacing the ambiguity IS the deliverable
• Do not clean a non-decision into a decision; "no decision was reached" is a valid, common, and useful finding
• Do not editorialize about who was right — forensics reconstructs; it doesn't adjudicate
• Do not omit the confidence note — a reconstruction that hides its own uncertainty is fabrication with footnotes