Global operating rules (all folders, all sessions)

Planning Doctrine — Single Round (2026-09-08)

Applies to every development request (platforms, systems, integrations, sites, scripts) that is more than a trivial single-file edit. Copy work and pure research follow their own pipelines but inherit the "no calendar" rule.

1. Plan first, wait for GO

Before writing code, produce a plan and stop for the user's GO. Trivial edits (one file, unambiguous, reversible) skip the plan.

2. Never plan in time

A plan NEVER contains days, weeks, hours, sprints, "round 1/2/3" as time, or phase durations. Work that takes minutes must not be dressed as a calendar. An external client deadline may appear once, as a constraint (a fact), never as a schedule for the work. Phases (A, B, C…) express dependency order only.

3. Plan structure (mandatory, 5 blocks) — template: ~/.claude/templates/plano-rodada.md

1. Objective + what already exists — one sentence, then a reuse table (what exists / what is missing).
2. Executable scope now — EVERYTHING that can be built with the information and access available, ordered by dependency. If the whole request is executable, the plan says "GO total". Nothing is deferred for lack of time, only for a real barrier.
3. Barrier — split in two: what blocks testing (what the human/client must do to validate) and what blocks continuing (missing information, access, decision). Every item becomes an explicit request. Decisions that are the user's to make are listed with a recommendation.
4. Done criteria + self-tests — what the agent verifies alone before reporting (build, unit/E2E tests, diff read-through, screenshot).
5. Usage estimate — table "How to execute" (rows = model routing options, columns = API-equiv cost, % Fable weekly, % general weekly), the recommended row marked, then one line of optimizations (which model at which step, subagents or not, minimal scope, reuse). Run claude-usage --estimate <usd> to convert cost into % and to print the current window state ("costs ~X% Fable · today at Y%, pace Z%").

4. Build to the barrier

After GO, execute the entire executable scope without stopping to ask, except for destructive or irreversible actions (production deploy, bulk sends, data deletion, campaign activation). Then run a review pass: check every plan item against what was built (tests, build, E2E, reading the diff). Any divergence becomes a report item, never silence.

5. Report structure (mandatory, 5 blocks) — template: ~/.claude/templates/report-rodada.md

1. Done vs planned — table: item → state → evidence.
2. Test script for the human/client — step by step, with expected results.
3. Requests for the next planning round — information, access, decisions.
4. Actual vs estimated usage — for calibration; note the model routing that was used.
5. Next barrier — what the next round will need.

6. Usage reference (calibrated 2026-09-02, Claude Max 20x)

Weekly caps ≈ $3,300 general / $1,770 Fable (API-equiv). Fable runs ≈ $65–110/h. Reference classes:

Routing defaults for the estimate table: Fable for architecture, hard bugs, final review; Opus for implementation volume; Sonnet for subagents and mechanical work. Always offer at least two rows (all-Fable vs. split routing) and mark the recommended one. The user decides the routing at GO.

Note: these weekly caps are calibrated for a Claude Max 20x plan. Each partner must recalibrate the CAP_ALL, CAP_FABLE and CAP_5H constants in claude-usage for their own plan before trusting the percentages.

7. Enforcement

A global PostToolUse hook (~/.claude/hooks/plan-no-time-lint.cjs) warns when a plan file (PLANO-*.md, *-plano*.md, brain/ops/plans/*.md) contains calendar patterns. Warning only — fix the plan when it fires.

Related: Agent System director doctrine at <repo>/.claude/agents/director.md.