version: "1.0.0" name: prime-context description: Use before starting pipeline work on a Ready feature the agent hasn't worked yet. Triggers — "/engineer.prime-context", "prime context", "load context before we start", "get up to speed on feature X".

prime-context

Load working memory before pipeline work on a Ready feature — the convergent counterpart to discuss. Forked from superpowers:brainstorming but inverted: no exploration, just loading and orienting. Produces no artifact.

When to use

A prep step (not a checkpoint) between feature-init and discover-acs, on any feature the agent hasn't already worked this session.

Skip when: the agent just created the feature this session (context already warm). Not for: exploring whether a feature is worth doing (discuss).

Workflow

1. Resolve + locate — resolve the methodology root + manifest via ${CLAUDE_PLUGIN_ROOT}/scripts/dae_resolve.py (see references/resolving.md); locate the feature (slug arg or branch name). Reject if no folder / no feature.md. Layout disambiguation: if the project has BOTH a features/NNN-*/ and a specs/NNN-*/ tree (a speckit migration in progress), state which holds this feature's artifacts; and if more than one acceptance-IR convention is present (features/NNN/.build/spec.json vs an acceptance-pipeline/ir/ per-GWT layout), name the one this feature uses — so the load doesn't rediscover the split ad-hoc. (speckit-consolidate unifies it.)
2. Silent batch load — without narrating: feature.md, CHARTER.md, manifest.yml, prior handoffs/ (especially the originating *-discuss.md), and the files named in feature.md's "Related code / design pointers". When loading those code pointers, prefer LSP — documentSymbol for file shape, hover for signatures/types, workspaceSymbol to pull in the symbols they reference — over reading whole files blind, when an LSP MCP capability is available; fall back to grep + Read otherwise. See ${CLAUDE_PLUGIN_ROOT}/references/code-lookup.md.
3. Orient — give a concise summary: outcome, scope, autonomy level (+ charter cap), key prior decisions, related code, relevant ADRs.
4. One prompt — ask exactly one question: anything else to load? If the user names a new code pointer, load it and offer to add it to feature.md. Then stop — prime-context orients, it does not interview.
5. Breadcrumb handoff — emit a tiny handoff recording what was loaded.

Re-invocation re-loads fresh (no incremental diffing).

Handoff

Emit per ${CLAUDE_PLUGIN_ROOT}/references/handoff-summary.md. checkpoint: null; artifacts: []; human_action_needed: no; recommended_next: "/engineer.discover-acs".

Always emit a breadcrumb. Earlier versions produced no handoff because prime-context creates no artifact. Users expect a breadcrumb after every skill invocation (distbute Jun 4: "emit the handoff") — surprising them with "no artifact by design" breaks the agentic-summary contract. The breadcrumb records: which feature was primed, what files were loaded, any new pointer added to feature.md, and the next recommended skill. Body can be one paragraph — frontmatter is the load-bearing part for progress-log propagation.

References

• Foundation Design — feature.md schema
• Discuss & Upstream Funnel — methodology_root resolution