`references/adoption-rubric.md`

113 lines
# Adoption rubric — should this tip become one of Steve's rules?

@tiktok_architect's feed is a firehose of *"never ask Claude Code to do X"* /
*"always do Y"*. The job is NOT to relay hype — it's to answer, per tip:
**codify, watch, or skip**, with a reason. Unlike a tool-scout, the thing being
adopted is a **behavioral rule** (a MEMORY.md feedback entry or a CLAUDE.md
line), so the first gate is truth.

## The CORRECTNESS GATE (runs first — a hard filter, not a score)

Before scoring anything, decide: **is this claim actually true for CURRENT
Claude Code / the current Claude models?** Verify with, in order of authority:

- the `claude-code-guide` subagent (features/hooks/CLI/MCP reality),
- the `claude-docs` skill (skills/frontmatter) and `claude-api` skill
  (models/pricing/tool-use), both `$0`,
- official docs — docs.claude.com, docs.anthropic.com,
- a second source via Exa/WebSearch if still unsure (metered — show cost).

Verdict of the gate:
- **FALSE / version-stale / can't be verified true** → **SKIP** immediately, with
  a one-line "why it's wrong" (and the correct behavior, if worth noting). Never
  score or codify an unverified behavioral claim.
- **TRUE (or true-with-a-caveat)** → proceed to the six-dimension score below.

This gate exists because short-form tips are engagement-optimized and routinely
outdated. A catchy wrong rule baked into MEMORY.md is worse than no rule.

## Dimensions (score each 0–3, 18 max — only for tips that PASS the gate)

1. **Correctness confidence (0–3)** — how solidly verified? 3 = confirmed by an
   authoritative source; 1 = plausible, thinly sourced; 0 shouldn't reach here
   (the gate caught it). A caveat-laden "true only sometimes" caps this at 2.
2. **Stack fit (0–3)** — does it apply to how Steve actually uses Claude Code?
   High: skills/subagents/CLAUDE.md/MEMORY.md/hooks discipline, permissions,
   cost/context management, catalog/scraper/DW pipelines, git hygiene. Low:
   generic influencer platitudes, things irrelevant to his workflow.
3. **Novelty vs. what Steve already codified (0–3)** — 3 = fills a real gap; 0 =
   already a rule in CLAUDE.md / MEMORY.md / an existing skill. **Always run the
   dedupe check below before scoring this.**
4. **Leverage (0–3)** — does following it save real hours / avoid a real failure
   / cut cost, or is it cargo-cult ritual? Favor "measurable" over "feels
   productive" (Steve's theater-vs-real test). A tip with a concrete *why* beats
   a bare assertion.
5. **Adoption cost / reversibility (0–3, higher = cheaper)** — 3 = a one-line
   MEMORY.md feedback entry, trivially reversible; 1 = needs a workflow change; 0
   = demands new paid tooling or a big rewrite to follow.
6. **Safety of ACTING on it (0–3, higher = safer)** — 3 = pure behavioral/local,
   no risk; 0 = the tip itself is dangerous ("disable all permission prompts",
   "let it auto-push", "run with --dangerously-skip"). A dangerous tip is not
   auto-DQ, but it's flagged loudly and the memo must warn against it.

**Verdict bands (after the gate passes):** ≥14 → **CODIFY** (draft the
MEMORY.md/CLAUDE.md memo). 9–13 → **WATCH** (log it; revisit if it recurs or a
second creator repeats it). ≤8 → **SKIP** (one-line why).

## Dedupe against what Steve already codified (before scoring novelty)

Steve's standing behavioral rules already live in `~/.claude/CLAUDE.md` and the
auto-memory. Check for an existing equivalent BEFORE crediting novelty:

```bash
# behavioral rules already in the global instructions
grep -in "<keyword>" ~/.claude/CLAUDE.md 2>/dev/null
# already-saved memories (index + bodies)
grep -rin "<keyword>" ~/.claude/projects/*/memory/ 2>/dev/null | head
# an existing skill that already encodes the practice
grep -ril "<capability>" ~/.claude/skills/*/SKILL.md 2>/dev/null | head
```

Examples of "he already has it": context/compaction discipline → CLAUDE.md
compaction rules + `token-use`; verify-before-claiming-done → `validation-loop`;
gated/irreversible actions → the whole yolo/officer gating model; cost visibility
→ the always-show-costs rule + `cost-tracker`; prompt-cache hygiene →
`prompt-cache-optimization`. If an equivalent exists, the useful output is a
**diff** ("the tip adds a specific threshold his rule doesn't state"), not a
redundant re-add.

## Gating — what the scout may and may NOT do

**Retrieval, transcript, verification, dedupe, scoring, and writing the
shortlist are reversible and run freely.** Everything that changes Steve's
standing configuration or spends is gated — draft it to
`~/.claude/yolo-queue/pending-approval/` with an APPROVE/REVISE/BLOCK block;
never auto-apply:

- **Editing `~/.claude/CLAUDE.md`** (his global standing instructions) — always
  gated. This is behavior-changing and irreversible-ish; Steve approves it.
- **Writing a new MEMORY.md feedback/user rule** derived from an external tip —
  gated. (The memory system lets Claude save memories, but adopting an outside
  creator's claim as a standing rule is a judgment change Steve signs off on.)
- Anything the tip tells you to *install/run* (a repo, an MCP, a CLI, a paid
  key) — gated; route the tool half through `gurisingh-scout`'s adoption path.
- Anything customer-facing / destructive / outbound the tip would trigger.

The deliverable is a **decision-ready shortlist of rules**, not edited
instructions. Adoption is Steve's call.

## Output shape (per run)

Produce a ranked table:
`Tip (claim) | Video (id, date) | Kind | Verified? (source) | Dedupe (have/gap) | Score | Verdict | 1-line reason`.

Then, for each **CODIFY**, a short pending-approval memo containing:
- the claim, restated as an imperative rule in Steve's voice,
- **why it's true** + the authoritative source that confirmed it,
- the exact proposed diff — either a `MEMORY.md` feedback entry (with frontmatter
  + **Why:** / **How to apply:** lines) or a specific CLAUDE.md line/section,
- adoption cost (`$0` to codify; note any tooling the tip implies), risks/caveats
  (loudly if the tip is dangerous), and an APPROVE/REVISE/BLOCK line.

Log genuine wins only — a tip Steve actually adopts as a rule, not the scan.