sleeper-fantasy-mcp-: Instructions file for Claude Code

Working rules for this project

The core rule: question every decision, then justify it

Do not make a choice and move on. For every decision — a formula, a threshold, a field, a default, a filter, a ranking — state what you chose, why, and what would have to be true for it to be wrong. If you cannot justify it, you have not finished thinking about it.

This applies to decisions that feel obvious. The expensive mistakes in this project all looked obvious at the time.

Concretely, before moving past a decision, ask:

• What does this field actually contain? Not what its name suggests.
• What would this look like if I were wrong? Then go check for that.
• Is this result plausible? If a number surprises you, chase it before accepting it.
• Does this hold for every configured league? They have different rules.
• What am I assuming that I have not verified?

Then say the answer out loud in the response. The reasoning is part of the deliverable, not scaffolding to be discarded.

Verify data semantics before modelling on a field

Every serious bug here came from trusting a field name.

• rec_air_yd looked like intended air yards. It is completed air yards — proven by rec_air_yd + rec_yar == rec_yd exactly. True aDOT is not derivable and the code says so rather than shipping a wrong number under a right name.
• rec_0_4, rec_10_19 etc. looked like target-depth buckets, which would have been ideal xFP inputs. They sum to receptions — they are outcomes, not opportunities, and would have leaked the answer into the model.
• Sleeper omits pts_ppr entirely for a player who dressed but scored nothing, rather than writing 0. That silently dropped 2,021 real zero-point weeks and inflated the floor of every low-usage player.
• Season stats return rows for seasons before a player entered the league, with 0 games. Uncorrected, every rookie looks maximally injury-prone.

The pattern: check that a field means what you think by testing an identity that must hold if it does.