| name | leveragepath |
| description | Identify the highest-leverage actions for the current product based on taste.md, taste.vision, and SPEC.md, with deepresearch-backed channel scanning, RICE+categorization scoring, auto-vs-manual classification, community targets with verified URLs, and (when research surfaces moats or blind spots) a /defineicp-style proposal to update the kernel. Use when the user invokes /leveragepath or asks "what's the highest-leverage move I can make right now?" |
| argument-hint | [{"mode":"scan|apply <move_id>|kernel-propose"}] |
| disable-model-invocation | true |
/leveragepath
Run the leverage-path workflow for:
$ARGUMENTS
/leveragepath finds the highest-leverage distribution, positioning, and
community moves available to the project right now, ranks them, separates
what Claude Code can run autonomously from what only the operator can do,
and — when deepresearch surfaces a moat or blind spot the dev was not
seeing — proposes a kernel update so the next iteration starts from the
new ground truth.
This command is not a marketing-checklist generator. It is a governed
research-then-route workflow: deepresearch first, ranked move synthesis
second, automation classification third, and kernel mutation only with
explicit kernel-propose opt-in and proposal/apply semantics.
For apply mode or any leverage-driven file mutation, /opusworkflow is the
default outer route and /leveragepath is the inner contract. Direct
/leveragepath invocation remains valid, but it must inherit the
Claude/Opus planner-reviewer plus MiniMax-M2.7-highspeed executor policy
before mutation.
outer_route: opusworkflow
inner_contract: leveragepath
When to use this
- The dev just shipped something and asks "where do I distribute?"
- The dev is stuck and asks "what's the highest-leverage move I can make
right now?"
- The dev suspects they have a moat they are not articulating publicly.
- The dev keeps doing the same medium-leverage moves and wants a fresh
surface scan.
- A community-led-growth or product-led-growth iteration is about to
begin and the dev wants the canonical channel list before committing.
Not for: writing the actual marketing copy (use the regular /workflow
or specialist channel tools), verifying that an executed move worked
(use /verify against SPEC.md), or generating a generic checklist
(this is research-first).
Inputs read at run start
taste.md, taste.vision — what the project values, who it serves,
what it refuses to do.
SPEC.md (active) — what is currently shipping or just shipped.
- Latest
.taste/workflow-runs/*-workflow.md — what the dev has already
attempted recently.
- Recent git log — what surfaces have been touched in the last week.
- Any
.taste/leveragepath/*.md from prior runs — moves already executed
or already declined, to avoid re-suggesting.
This skill does NOT read .env, .env.*, .claude/settings.local.json,
secrets/**, customer artifacts, or private connector files.
Modes
scan (default)
Run the full leverage-path identification pipeline:
- Read inputs above.
- Define the project category (open-source dev tool? B2B SaaS? indie
game? academic project?) — used to scope which channel surfaces apply.
- Invoke
/deepresearch with effective parallel budget on these
branches (use only the lanes that materially help):
- Communities: which Discord servers, Slack workspaces, subreddits,
forums, mailing lists are active in this category in 2026?
- Marketplaces / directories: which awesome-lists, plugin
directories, app stores, package registries are canonical?
- Influencers / curators: who covers this category on blogs,
newsletters, podcasts, YouTube?
- Conferences / events / live windows: anything time-sensitive
in the next 30 days?
- Adjacent tools / collaborators: who is building close-but-not-
competing? PR / mention / collab opportunities?
- Academic surfaces: arXiv, conference workshops, paper-citation
lanes if the project has academic legitimacy.
- For each move identified, compute a RICE-style score:
- Reach — estimated audience size of the channel (verified live).
- Impact — expected effect on the operator's stated goal (signups,
stars, paying customers, citations, etc.) on a 0.25–3 scale.
- Confidence — how sure we are the move will land (60% / 80% /
100%).
- Effort — total operator + Claude Code time in person-hours.
- Score = (R × I × C) / E.
- Add four orthogonal tags per move:
auto (Claude Code can run it via tools) | manual (operator only)
community (targeted) | mass (broad)
time-sensitive (window in next 30d) | evergreen
reversible (can undo in <1 day) | hard-to-reverse (public commit
to a brand or relationship)
- Run
/introspect pre-write gate — challenge weak sources, missing
surfaces, over-confidence in unverified channel sizes, missed
blind spots.
- Write the artifact to
.taste/leveragepath/<run_id>/leveragepath.md with sections:
- Top 5 moves by RICE (the leverage punch list)
- Auto-by-Claude Code (with the exact tool calls it should run)
- Manual-by-operator (with copy-paste assets where applicable)
- Time-sensitive (next 30 days) (windows ranked by start date)
- Communities to target (Discord/Slack/forum invite URLs verified
live this run)
- Adjacent collaborators (named, with public outreach surface)
- Moats observed (research-surfaced advantages the dev is not
publicly articulating)
- Blind spots observed (channels or angles the dev has not
considered, ranked by leverage)
- Source ledger (cited / reviewed / rejected, with access date)
- Kernel mutation candidates (if any — proposal-only, never
applied in scan mode)
apply <move_id>
Execute one move from the most recent scan artifact. Only auto-tagged
moves are eligible for apply; manual moves return a "blocked: operator
must execute" status.
Auto moves typically include:
- Open a GitHub issue or PR via
gh CLI.
- Submit a directory-listing PR to an awesome-list.
- Update a static site with a positioning shift.
- Write and queue a draft email via
mailto: link generation.
- Pin GitHub repos via API (where the operator has token).
apply mode runs the standard /opusworkflow outer route inside it:
SPEC fragment for the move, /specqa, execute, verify against the
move's success criteria from the scan artifact, post-execution
/introspect. Closeout records the move's outcome back to the scan
artifact for future iterations to skip.
kernel-propose
Read the most recent scan artifact. Extract the Moats observed and
Blind spots observed sections. For each, produce a proposed diff to
taste.md and/or taste.vision that surfaces the moat in the project's
public positioning OR closes the blind spot in the project's strategic
view.
kernel-propose produces a proposal artifact only. It does NOT modify
taste.md or taste.vision. Apply uses the standard /defineicp
proposal/apply semantics: explicit operator approval, file backup, hash
record, changed-line trace, validation, rollback evidence.
Required scoring shape
Every move in the artifact must include:
### Move N: <short name>
| Field | Value |
|---|---|
| RICE | (R × I × C) / E = <score> |
| R (Reach) | <number with unit and live source> |
| I (Impact) | <0.25 / 0.5 / 1 / 2 / 3> with one-line rationale |
| C (Confidence) | <0.6 / 0.8 / 1.0> with one-line rationale |
| E (Effort) | <person-hours, sum operator + Claude Code time> |
| Auto/Manual | <auto with tool list, OR manual with reason> |
| Community/Mass | <community: target> OR mass |
| Time | <evergreen, OR sensitive: window start–end dates> |
| Reversibility | <reversible, OR hard-to-reverse with reason> |
| Blocker if any | <none, OR explicit blocker> |
| Next concrete step | <single command, URL, or copy-paste asset> |
The Top 5 list orders by descending RICE.
RICE framework lineage
RICE (Reach × Impact × Confidence / Effort) is the canonical product
and growth prioritization framework as of 2026. Sources verified live:
This skill uses RICE as one layer in a four-layer stack:
- Layer 1: RICE numeric ranking.
- Layer 2: auto/manual orthogonal tag.
- Layer 3: time-sensitivity orthogonal tag.
- Layer 4: reversibility orthogonal tag.
The numeric ranking decides ordering; the orthogonal tags decide which
moves the dev should execute first this hour vs this week vs this month.
Sibling-skill integration
/deepresearch is the research engine. /leveragepath scan invokes it
internally with the effective parallel budget computed via
bash scripts/parallel-capacity.sh --json.
/introspect runs as the pre-write gate before the artifact is
written. Same hard-block semantics as the rest of the harness.
/defineicp is the kernel-mutation contract for kernel-propose
mode. The proposal artifact uses the same shape: target file backup,
hash, changed-line trace, validation, rollback.
/opusworkflow is the outer route when apply <move_id> mutates
files. Records inner_contract=leveragepath in the run artifact.
/icpweek is a sibling product-stress lens, not a substitute. ICP
week-in-the-life surfaces product mismatches with users; leveragepath
surfaces channel mismatches with reach.
/visualizeworkflow --continue may follow a leveragepath scan when
the dev wants a visual of the channel surface before committing to
any move.
Hard rules (no exceptions)
- Every channel reach number must be verified live this run (citation
with access date). No pretrained-memory-only audience sizes.
- Every community URL must be verified to resolve to an active
surface this run. Dead/stale links flagged or removed.
- Move rankings must show R/I/C/E components, not just the score, so
the operator can challenge the inputs.
apply mode must not execute a move whose Auto/Manual field is
manual — it returns "blocked: operator action required" instead.
kernel-propose mode must not write to taste.md or taste.vision
— the proposal is artifact-only; mutation requires a separate
operator-triggered /defineicp apply invocation.
- The skill must surface at least one moat and at least one blind
spot in scan mode, OR explicitly state "no moats / blind spots
surfaced this run" with the introspection rationale. Silence is
not acceptable.
- The artifact source ledger must list at least three distinct sources
and may not consist entirely of the project's own surfaces.
Anti-patterns
- One-shot search-and-summary masquerading as a leverage scan.
- Ranking only by reach (ignoring effort and confidence).
- Mixing automatable and manual moves in a single execute command.
- Recommending a time-sensitive move whose window already closed.
- Skipping
/introspect because the channel list "looks complete."
- Proposing kernel changes in scan mode (kernel-propose is opt-in).
- Failing to cite live URLs for community / directory targets.
- Promising linear scaling: "post on N channels = N× the leverage" — same
problem honest-eta blocks for time estimates.
Output artifact contract
.taste/leveragepath/<run_id>/
leveragepath.md (the human-readable artifact)
leveragepath.json (the machine-readable summary, if scripts
are added in v2; v1 ships markdown only)
source-ledger.md (cited / reviewed / rejected sources;
optional — may be inlined in leveragepath.md
for v1)
Run IDs use the convention YYYYMMDD-HHMMSS UTC, matching other
.taste/ subtree run IDs.
Provenance
This skill exists because in the 2026-05-11 session that produced the
waitdeadai/llm-dark-patterns distribution wave, the operator and Claude
Code organically performed all of these steps in an ad-hoc manner —
identifying the Anthropic Discord (95k members) only after explicit
research, sorting moves into auto vs manual under live time pressure,
and missing several time-sensitive windows because no canonical surface
scan ran first. /leveragepath codifies that ad-hoc work into a
governed, repeatable skill.