tl

Tackline

Composable workflows for agent-driven development. Clone once, symlink into ~/.claude/, use across every project.

In naval semaphore, a tackline separates signal groups to prevent ambiguity. Here, it separates composed skill outputs -- each skill's result feeds the next through conversation context, like Unix pipes for knowledge work.

What It Looks Like

Map an unfamiliar codebase

/blossom audit the event sourcing pipeline for consistency gaps

Blossom spawns parallel spike agents that read actual source files, trace call chains, and tag findings with confidence levels. Output is a prioritized task backlog, not a summary.

Seeding epic: "Event sourcing pipeline audit"
Dispatching 4 spikes: command-handlers, projections, event-store, saga-orchestration

Spike results:
1. CONFIRMED -- ProjectionRebuilder skips tombstoned events (projections/rebuild.ts:47)
2. CONFIRMED -- No idempotency check on CommandBus.dispatch (handlers/bus.ts:112)
3. LIKELY    -- Saga timeout defaults to 0, may cause silent drops (sagas/orchestrator.ts:88)
4. POSSIBLE  -- Event schema v2 migration has unused backward-compat shim

Created 6 tasks under epic BL-42, dependency graph attached

Hash out a design decision

/meeting should we use event sourcing or CQRS-lite for the new billing module?

Assembles 2 panelists with genuinely opposed perspectives and a facilitator. You steer the conversation, ask follow-ups, cut threads that aren't productive. Output is a decision record, not a compromise.

Research, distill, and prioritize

/gather authentication patterns in this codebase
/distill
/rank by security risk

Each skill's output feeds the next through conversation context. No file passing, no explicit piping -- context is the pipe.

Install

# Add the marketplace (one-time), then install
claude plugin marketplace add tyevans/tackline
claude plugin install tackline@tackline

# Install global rules (plugin system doesn't support rules yet)
# No local checkout needed — fetch directly from GitHub:
curl -fsSL https://raw.githubusercontent.com/tyevans/tackline/main/dev/install-rules.sh | bash
# Or, from a local checkout:
# bash /path/to/tackline/dev/install-rules.sh

OpenCode

Tell OpenCode to fetch and follow the install guide:

Fetch https://raw.githubusercontent.com/tyevans/tackline/main/docs/install-opencode.md and follow the instructions to install tackline.

See docs/install-opencode.md for details.

Uninstall

claude plugin uninstall tackline@tackline
# Remove rule symlinks (leaves non-tackline rules intact)
for f in /path/to/tackline/rules/*.md; do rm -f ~/.claude/rules/"$(basename "$f")"; done

Getting Started

New project: Run the project-bootstrapper agent to set up CLAUDE.md, hooks, and rules. Then run agent-generator to create project-specific agents. Start exploring with /blossom.

Existing project: Type /blossom <what you want to explore> to map the territory. Use /consolidate when the backlog grows noisy. Use /session-health for a context quality gut-check.

New machine: Clone, install, go. Your entire Claude Code workflow travels with you.

Skills (48)

14 composable primitives -- stateless skills that chain through conversation context (/gather, /distill, /rank, /filter, /assess, /verify, /critique, /diff-ideas, /decompose, /plan, /expand, /transform, /sketch, /merge).

25 workflows -- orchestrated multi-step workflows with side effects (/blossom, /fractal, /meeting, /consensus, /tracer, /review, /sprint, /spec, /premortem, /bootstrap, /drive, /deploy, /optimize, /test-strategy, /evolution, /drift, /consolidate, /bug, /domain, /active-learn, /diagnose-agent, /challenge-gen, /challenge-run, /storm-prep, /formalize).

9 session & team skills -- /status, /advise, /assemble, /standup, /retro, /handoff, /session-health, /team-meeting, /discover.

Full catalog with decision tree and chain patterns

Common Chains

/gather -> /distill -> /rank       Research -> condense -> prioritize
/decompose -> /plan -> /spec       Break down -> sequence -> specify
/blossom -> /drive                 Explore -> autonomously implement

All chain patterns

How It Actually Works

Skills are LLM-driven, not deterministic. The same input may produce different findings across runs. Confidence levels (CONFIRMED, LIKELY, POSSIBLE) help you calibrate trust. Output quality scales directly with input specificity -- a precise goal with file paths produces better results than a vague one.