Outline Code

outline-code

Produce a flow document for a code scope: every function traced as an indented outline where each indent level is a logic fork, every node cites file:line, plus the implicit coupling between flows (channels, watches, queues, shared state) as a wiring table. The document exists so a reader can hold the whole subsystem's control flow in their head and act on a goal — usually spotting ways to condense the logic.

Phase 0 — Scope, goal, output

Parse these from the prompt before reading any code:

• Scope: the @-mentioned files, named module/crate, or directory. Then expand to coupled context — the document is far more useful with it: (a) the spawn/call sites of the scope's entry points, (b) the upstream writers that feed any channel/watch/shared state the scope reads, (c) the downstream consumers of what the scope produces. The scope's own files get full fork-level detail; coupled files get traced only along the paths that touch the scope.
• Goal: what the user wants the document for. "Condense / simplify / loosely organized" → condensation opportunities. "Find bugs / suspicious" → risk observations. "Onboarding / understand" → reading-order notes. No stated goal → default to condensation opportunities. The goal shapes only the final analysis section; the outlines are always neutral.
• Output path: explicit path if given; otherwise -flow.md at the repo root.

Ask the user only when something is genuinely undecidable — two plausible scopes, a goal that could be read two ways, or a presentation choice the prompt doesn't settle (e.g., mermaid vs tables). Otherwise infer and proceed; state your inferences in the final summary so they can be corrected.

Phase 1 — Trace the source map first

Before writing anything, build a map by grepping, not by reading files end-to-end:

1. List the scope's functions and entry points (spawned tasks, public API, handlers).
2. For each entry point, find its call/spawn sites — including outside the scope. Note the gating conditions (mode checks, feature flags, config).
3. Find the coupling mechanisms: channel/watch/queue accessors, shared structs, event topics, globals. Grep each name across the repo to enumerate every producer and consumer site. Whatever the codebase's mechanism is — a message bus, tokio watches, event emitters, pub/sub, callbacks — that's the wiring table's subject.
4. Record everything as item → defined at → called/spawned from with file:line.

This map is the input for every later step. Getting it right up front prevents agents (or you) from mislabeling producers/consumers later.

Document structure

Use this exact skeleton:

1. Title + notation line — one sentence stating what the doc traces, then: outline notation is - event/condition → function_name (file.rs:line); each indent level is a logic fork. Follow with a file key table mapping the bare file names used in citations to full paths.
2. Overview — the long-running tasks / entry points as a list, each with function (file:line), its spawn site, and the condition gating it (mode, flag). End with one paragraph giving the mental model — typically who produces targets, who produces data, who consumes — refined to match what the code actually does.
3. Coupling wiring table — one row per channel/shared-state item: name | primitive + definition line | producers (each function (file:line)) | consumers (same). After the table, at most 2-3 one-line notes for non-obvious semantics (dedupe-before-send guards, mutex-wrapped receivers, lookalike fields that are NOT part of the wiring).
4. Flow outlines (h2) — one h3 per flow. Order: startup topology first, then one section per major flow following the data (e.g., trigger → transform → consume), the downstream consumer last. If verification finds functions no flow covers, add a final "helpers called from outside the flows above" subsection rather than silently dropping them — the doc's claim is every function.
5. Analysis section (goal-adapted; default heading "Condensation opportunities") — titled bullets, 2-4 sentences each: what's duplicated/awkward with file:line for each occurrence, then the concrete remedy (helper name/shape, merged function, enum). Order by impact. Mark borderline items lower-confidence and say what would need confirming. No code blocks unless a 2-3 line sketch genuinely clarifies.

Outline style contract

These rules are what make the document scannable; hold every section to them:

• Node format: - → function_name (file.rs:line). Condition first, then arrow, then target. Branches that don't call a function: - → (file:line).
• 2-space indent per level. Indentation means exactly one thing: a logic fork. Every if/else, match arm, select! arm, early return, and loop exit in a traced function appears as a child node under the node where the fork happens.
• One line per node. No prose paragraphs inside outlines. A single one-line lead-in sentence per section is allowed.
• When a flow reaches another flow's territory, reference it (→ epoch pack download flow) instead of expanding it — each flow is documented once.
• Overloaded semantics deserve a callout on the node itself (e.g., a return value meaning retry vs continue vs done).
• Completeness bar: every non-test function in the scope appears at least once; every branch point appears as a fork. Coupled files only need the branches on paths that touch the scope.

Example (from a consensus state-sync trace):

- watch updated → `spawn_track_recent_consensus` loop wakes (consensus.rs:130)
  - `rx_gossip_update.changed()` fires → forward `(epoch, number, hash)` into `consensus_request_queue` (consensus.rs:142-145)
  - `rx_shutdown` fires → return (consensus.rs:148-150)
- queue item received → `manage_new_consensus` (consensus.rs:314)
  - first gossip: `first_gossipped_epoch.is_none()` → spawn backtracking fetch to genesis (consensus.rs:332-354)
  - subsequent gossip → throttle on AtomicI32 `tasks` (consensus.rs:355-359)
    - `task_num = 6` → skip; subsumed by later gossip (consensus.rs:360)

Execution — scale to the scope

Small scope (roughly ≤2 files / ≤800 lines of in-scope code): trace and write inline. Still do Phase 1 first and still run the verification pass below.

Larger scope: parallelize with subagents, one agent per document section — never bundle sections into one agent. Spawn the whole batch in one message: one agent each for the overview, the wiring table, and each flow outline. Give every agent:

• the exact file paths and relevant line ranges,
• the Phase 1 source map (as hints to verify, not trust),
• the style contract above,
• an output contract: "Your final message is consumed by an orchestrator assembling a larger document. Return ONLY the finished markdown for your section, starting with its heading — no preamble, no commentary. Verify every line number against the source you read."

Then assemble and normalize. Expect cross-agent discrepancies — two agents will describe the same producer/consumer with different file or function labels. Resolve every conflict by reading the cited source yourself before assembly; never pick one agent's label over another's without checking the code.

After assembly, spawn one analysis agent for the final section. It reads the assembled document plus the source, so its observations build on the verified outlines. Seed it with any duplication/awkwardness you noticed during Phase 1 and assembly, and tell it to confirm or drop each seed with line evidence and to find what the seeds miss.

Verification — always, before presenting

Two independent checks (parallel agents for large scopes, inline for small):

1. Citation check: extract every file:line reference and verify the named construct sits there (±2 lines). Read each source file once and batch-check its citations. Report wrong vs off-by-small separately.
2. Completeness check: (a) enumerate every non-test function in scope — each must appear in the doc by name; (b) walk every traced function's branch points — each must appear as an outline fork; (c) grep every coupling accessor — each non-test usage site must appear in the wiring table.

Fix everything found: correct off-by-small anchors, fix mislabeled function names in the table, add missing wiring rows, and cover missed functions (usually as the "helpers" subsection). This pass routinely catches real errors — in practice it has found missing functions, a missing table consumer, and wrong function labels that section agents introduced confidently.

Finish by summarizing for the user: what was produced, the inferences made in Phase 0, the verification results (citations checked, anything fixed), and the top findings from the analysis section.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: grantkee
• Source: grantkee/claude-extensions
• License: MIT

Install and usage instructions live in the source repository linked above.