report

Writing the session up

One session in, one report out: argued from what the session actually established, designed around its own subject, written to disk as a page you can read and a PDF you can send, and offered in three registers so the same evidence reaches a specialist, a decision-maker and someone meeting the subject for the first time.

Using Gemini? Read gemini.md for this skill's task-specific calibration and artifact checks. Its older runs do not establish Gemini 3.8 capability limits: preserve the user's selected implementation lane, and apply a model-specific route only when the current model or an observed failure supports it.

The failure this exists to prevent

A session reads forty files, runs a dozen commands, and reaches a conclusion. Asked for a write-up, the obvious move is to narrate the conclusion in confident prose. The result reads well and is unusable, because three very different things render identically:

• "the queue drops 12% of events" — measured, from a command whose output is in the transcript
• "the queue drops 12% of events" — read off one log sample and generalised
• "the queue drops 12% of events" — inferred from two facts that were each established separately

A week later nobody can tell which. Neither can the person you sent it to, and they have less reason to trust you than you do.

So the spine of this skill is: compile the ledger before you design anything. Every claim gets a locator or a label. The citation UI is then generated from the ledger rather than retrofitted onto finished prose, which is the only way the two stay honest with each other.

references/evidence.md carries the research behind every rule below. Read it when you need to justify or tune a rule, not on every run.

Resolve dependencies in the current runtime. Use the exact installed identifiers for the named plugin skills. dataviz is a standalone skill identifier only when the runtime lists it; if unavailable, use this skill's references/visualisation.md and an available charting or diagram tool, and name the substitution. Discover media and Mobbin tool names from the tool catalog. Reuse a loaded skill's guidance and the same acceptance evidence across phases; routing every decision through its criteria does not require invoking the skill again for each word or figure.