Prompt Coach — Configuration & Dashboard

ack_clean

bool

true

Emit a compact ambient one-liner on a clean prompt (nothing else fired) confirming the coach ran + showing the active rule closest to mastery. A liveness heartbeat, not praise — distinct ✓ glyph, informational not evaluative. On by default; set false to go silent on clean prompts.

ack_ratio

int

1

Rate-limit for ack_clean: emit the liveness line every Nth clean prompt. 1 = every clean prompt (default; a steady heartbeat). Raise to 5/10 for a quieter pulse.

collaborator_gate

bool

false

Make the collaborator block a real confirmation gate. Default false = honest proceed: the block says it’s proceeding (not 'reply yes to proceed'), Claude answers the same turn from the rewrite, and states which prompt it used (silence = accept; 'no'/'edit' next turn corrects). Set true to make Claude STOP after the block and wait for your yes/no/edit before doing the work — full control, at one extra round-trip per fired prompt.

enabled

bool

true

Master switch (v0.29+). When false, the coach’s UserPromptSubmit hook returns immediately without analyzing, logging, or emitting. Use pause_until_prompt for temporary silence; use enabled=false to fully disable the coach for this scope. Rendering is always inline as of v0.29 — the pre-v0.29 nudge_style options (both/silent/log-only) are silently ignored.

library_hints

bool

true

When a rule fires, ground the collaborator rewrite in the closest gold-standard template from Anthropic’s Claude Code Prompt Library (a vendored, offline snapshot). On by default; the hint is only added when a confident task match exists. Set false to keep rewrites un-anchored.

max_nudges_per_window

int

6

FATIGUE CAP (v0.41+): max visible rewrites within nudge_window prompts; over the cap, fires are still logged and bookkept but the rewrite isn’t rendered. 0 disables.

nudge_window

int

20

Rolling window (in prompts) for the fatigue cap (v0.41+).

pause_until_prompt

int

0

Skip all nudging until global prompt_count exceeds this number. Use to silence the coach for N prompts (say 'coach pause N').

show_source_urls

bool

true

Show full clickable doc URLs in the coach’s Sources line (the collaborator block), so you can Cmd/Ctrl-click to open the Anthropic guide section. On by default. Set false to show just the short anchor slug (e.g. 'be-clear-and-direct') and keep the block compact.

tip_cooldown_prompts

int

100

Minimum prompts between two fires of the same tip (v0.28+). Anti-nagging cap; the coach shouldn’t remind you about the same technique every few prompts.

tip_ratio

int

5

Variable-ratio: 1 in N matching + cooldown-clear opportunities actually fires a tip (v0.28+). Lower = more frequent tips.

tips_enabled

bool

true

Enable proactive tips (v0.28+): 💡 suggestions pointing at advanced techniques you could try. Distinct from rules which fire on prompt problems — tips fire on-topic for techniques you haven’t used. Fires on matching prompts (rate-limited) and on rule masteries (paired scaffolding).

cooldown_prompts

int

5

Minimum prompts between two fires of the same practicing rule (anti-nagging cap).

disabled_rules

list[str]

[]

Rule ids to permanently silence. Say 'coach off <rule-id>' to append or 'coach on <rule-id>' to remove.

explore_period

int

10

Explore/exploit (v0.41+): re-admit one dormant rule every N prompts to refresh its acceptance estimate so a rule noisy in only one context isn’t buried forever. 0 disables.

graduation_threshold

int

15

Clean prompts in a row a rule tracks as a recency/decay signal. Since v0.40 this no longer drives mastery (see min_demonstrations); it feeds the regression guard and the inactive-after default.

inactive_after

int

15

Clean_streak with ZERO demonstrations after which a rule retires inactive ('N/A to how you work') instead of lingering as practicing (v0.40+). Defaults to graduation_threshold.

max_active_rules

int

6

Cap on practicing rules active at once. As lower-tier rules master, higher-tier ones activate up to this cap.

min_demonstrations

int

3

EARNED MASTERY (v0.40+). Number of times a rule’s mirroring positive detector must fire — i.e. times you actively USED the good technique — before the rule can master. Absence of the mistake no longer counts; demonstration does.

min_fires_for_mastery

int

1

LEGACY (v0.27) evidence gate, superseded by min_demonstrations in v0.40. Retained for forward-compat; ignored by the demonstration-driven graduation path.

min_outcomes_for_gating

int

4

How many recorded accept/edit/reject outcomes a rule needs before the precision gate applies to it (v0.41+).

precision_floor

float

0.15

Acceptance-rate floor (v0.41+): a rule below this over min_outcomes_for_gating recorded outcomes is demoted dormant. Default 0.15 (only rules rejected ~85%+ of the time).

precision_gating

bool

true

ADAPTIVE ACTIVATION (v0.41+). When on, a rule whose acceptance rate (accepted+edited)/outcomes falls below precision_floor is demoted to dormant and stops firing; an explore slot periodically re-surfaces it.

record_silence_as_accept

bool

true

In honest mode (collaborator_gate=false), if a collaborator rewrite was rendered last turn and you move on without objecting, record that as an implicit acceptance (v0.49+). Populates the acceptance ledger so precision-gating has a denominator; without it the ledger stays near-empty and gating never activates. Explicit 'no'/'redo' still counts as a rejection. Ignored in gate mode.

regression_guard

int

3

Clean prompts since the last fire required alongside the demonstrations for mastery (v0.40+) — proves no active relapse at the moment of graduation.

review_intervals_days

list

[30,90,180]

DECAYING MASTERY (v0.41+): expanding review schedule in days of NON-USE. A mastered rule unused past the current interval decays to watch and must be re-demonstrated; each natural use resets + advances to the next interval.

demote_on_regression

obj

{"enabled":false,"threshold":3,"window":30}

Auto-demote a mastered rule that fires threshold+ times within window prompts. Shape: {enabled, threshold, window}. Off by default — surprise reactivation feels punitive.

mastered_cooldown_prompts

int

50

Cooldown between refresher fires on mastered rules. 10x the practicing cooldown by default. Set to 0 to disable refresher firing (permanent silence on mastered rules).

disable_praise

bool

false

Silence all praise but keep nudges. Praise+correction on the same prompt would dilute both (Kohn).

praise_novelty_window

int

5

Don’t repeat the same praise phrasing within N praises.

praise_on_first_after_fire

bool

true

Celebrate when you correct the exact thing you were nudged on in the previous prompt.

praise_on_mastery

bool

true

Celebrate whenever a rule graduates to mastered.

praise_ratio

int

10

1 praise per N clean prompts with a positive fire (variable-ratio, Kohn’s don’t-dilute threshold). Lower = more frequent.

typo_tolerance

int

2

Levenshtein distance for typo normalization (0 disables). Adaptive: distance-1 for short tokens (≤6 chars), distance-2 for longer.

llm_fallback

obj

{"enabled":false,"model":null,"min_words":20}

Opt-in stub. If enabled AND no rule fired via regex AND the prompt is long, an optional model call would classify against the rule catalog. Deferred until real-use data justifies it.