Claude Code Skill

Plan before you burn.
worktrickle

View on GitHub
claude — worktrickle demo

Token burn — same task

naive multi-agent
0
worktrickle
0

Workflow tool typically spawns 6–10 full-context agents with unconstrained outputs — several hundred thousand tokens before synthesis.

One knob to rule them all.

The effort dial scales fanout, agent caps, output caps, model tier, and circuit breaker settings together. Default: high.

Agent caps
8 concurrent / 24 total
Scout output cap
1500 tok
Worker output cap
3000 tok
Verifier output cap
2000 tok
Model tier
haiku scouts + opus implementors
Circuit breaker
pause at 4x est

Seven phases. Zero wasted agents.

Every run follows a canonical pipeline. Trivial tasks skip the pipeline entirely. The approved diagram is the run's contract.

01 · Triage
Inline. Decides whether to pipeline at all. Trivial tasks execute immediately with no fanout — the cheapest agent is the one never spawned.
inline · session model · ~1k
02 · Scout
One read-only haiku agent maps the codebase. Finds call sites, identifies partitions, surfaces viable strategies. Strict output cap: 400–3000 tok depending on effort level.
haiku · read-only · output-capped
03 · Plan
Inline synthesis of scout findings into a concrete execution plan: partitions, wave ordering, model assignment, token estimates per step.
inline · ~3k
04 · Diagram + Approve
Renders the ASCII workflow diagram — nodes, edges, model tiers, token estimates, dollar budget — and blocks until you approve, edit, or cancel. The diagram is the contract.
inline · approval gate · non-skippable
05 · Execute
Wave-parallel workers with file-disjoint partitions. Prompt-cache-aware: byte-identical preamble, sibling spawns batched inside the 5-minute TTL. Each worker reports DONE|PARTIAL|FAIL in a rigid grammar.
haiku/sonnet workers · cache-warm batching
06 · Verify
One sonnet verifier sees only the diff and acceptance criteria — not the full codebase. Reports correctness-affecting findings only. Clean: "PASS."
sonnet verifier · diff-only scope
07 · Synthesize
Inline ledger: est vs actual per step, total token and dollar spend, final summary. Writes structured state to /tmp scratch for resumability.
inline · ledger + state file

Right model, right role.

Haiku scouts at ~1/3 Sonnet cost. 40–85% of total savings come almost entirely from this one choice.

read-only scouting
haiku
~1/3 Sonnet cost. Maps code, finds call sites, identifies strategies. Strictly read-only, hard output cap.
mechanical workers
haiku
Repetitive substitutions, format transforms, boilerplate. No judgment needed — haiku suffices.
implementation
opus 4.8
Judgment writes. The default implementor at high+; sonnet handles routine items.
verification
sonnet
Correctness-only review of diff + criteria. Sees no more context than required. PASS or findings.
high-stakes arbitration
fable
At most one direct API call per run (~$0.07). Plan arbitration or final design adjudication. Degrades gracefully without API key.
triage / plan / synth
inline
The orchestrating session model handles planning, diagram rendering, and synthesis — no agent spawn cost.

Install in thirty seconds.

No runtime, no daemon. A markdown playbook that makes the main Claude Code session the orchestrator.

curl -fsSL https://raw.githubusercontent.com/ez-gz/worktrickle/main/install.sh | bash

Or clone manually:

git clone https://github.com/ez-gz/worktrickle && ln -s "$(pwd)/worktrickle/worktrickle" ~/.claude/skills/worktrickle

Optional — enable Fable escalation: export ANTHROPIC_API_KEY=sk-ant-...
Hooked on Fable but not on its API bill? worktrickle buys its judgment by the decision (~$0.07), not by the session.
Without a key, worktrickle degrades gracefully and never proposes Fable.

Full docs and source: github.com/ez-gz/worktrickle