Agent Reference Workflows

An index of repositories that each demonstrate one agentic workflow, end to end.

These are not tutorials and not templates. Each one is a record of an agentic workflow actually being run on a real task — the sequence that worked, the wrong turns, and the places the process fell short.

Why keep these as an index

I write a lot of Claude Code plugins. The hard part of designing a plugin is almost never the mechanics; it is knowing what the workflow should actually be — which steps exist, what order they go in, where a human has to intervene, and which step quietly does most of the work.

That knowledge only comes from having run the workflow properly at least once, in the open, with the mess left in. So these repos are kept as working context for plugin design. Specifically, each one is meant to supply:

The intended workflow	The stages, in order, as they turned out to be — not as they were first imagined
Where it broke	The steps that ran too late, the constraints that turned out not to discriminate, the assumptions that survived too long
Where a human is still required	Which is usually the most interesting line in the whole repo, because it is the line that moves
Candidate automation	The steps that are mechanical enough to become a command, a skill, or an agent

A plugin designed from a tidy imagined workflow encodes the imagined version. A plugin designed from one of these encodes what the work is actually like.

The index

Purchasing and evaluation

Agentic-Purchase-Pipeline An eight-stage process for evaluating any purchase over roughly $50, with a full worked example — a hardware purchase researched, specified, reversed four times, and ordered.

• Demonstrates: spec capture with evidence tagging · identifying the load-bearing requirement no vendor publishes · classifying a market by failure mode · driving a logged-in storefront to confirm variant-level availability
• Exposes: the spec-revision loop is the whole game and most tooling omits it; landed cost was computed one stage too late and overturned the recommendation at the last step; execution is still manual for authorisation reasons rather than capability ones

Development process

Research-Decide-Implement A deterministic Research → Decide → Implement pattern that deliberately overrides the default collaborative loop of coding agents: stay on the defined path and no approval gates are needed.

• Demonstrates: trading interactivity for autonomy under an explicit contract; artifacts as the communication channel between phases
• Companion: rdi-template — the canonical seed scaffold planted into a target repo

Document-As-You-Go A system prompt and slash command that make coding agents capture what they discovered, not just what they changed.

• Demonstrates: treating knowledge as an output of the work rather than a byproduct; routing a finding by its reach — this task, this project, or a standalone knowledge repo
• Exposes: capture reliably loses to task completion unless it is made an explicit obligation before reporting done

Decision-making under a spec

Spec-Led-Certification Choosing a professional certification spec-first: write the requirements, then survey and score the options against them.

• Demonstrates: the same shape as the purchase pipeline applied to a decision with no product to buy — evidence that the pattern generalises past shopping
• Related: Spec-Standard-Research — background research into specification standards for spec-driven agentic work

Reasoning and research

Perspective-Finder-Prompt A prompt for mapping the real nuances in a debate rather than flattening it to two sides.

• Demonstrates: structured disagreement-mapping as a repeatable agent task

Not currently published

Listed because they are part of the same collection and are relevant context, even though the repositories are private or local-only:

Repo	Workflow	Status
Claude-Shopping-Eval-Demo	Screenshot store catalogues → multimodal agent → ranked PDF. Complements the purchase pipeline: better when the constraint is price across local stores rather than fit	Local only — remote 404s
Ezra-Travel-Bot	Proof of concept for a flight-booking agent — the same execution boundary problem as stage 6 of the purchase pipeline	Private
Claude-Geolocation-Helper	Claude Code pattern for assisting with geolocation tasks	Private
Scraping-Notes	Verified web-access routes per site — what worked, what did not, and why	Private
Habits-Of-Claude	Standing habits for a coding agent, one file per habit	Private

What qualifies for this index

A repo belongs here if it:

1. Documents one workflow, not a grab-bag of prompts
2. Was actually run on a real task, with a worked example attached
3. Records where it went wrong, not only where it went right
4. Names the point at which a human is still required

A repo that is only a prompt, only a template, or only a finished result belongs elsewhere — see Claude-Code-Repos-Index for the full collection, and Subindices for the other indexes.

Deliberate overlap with the domain indexes

This index cuts by artifact type — "a worked run of a workflow" — while the domain indexes cut by subject. So entries appear in both, on purpose.

Most relevant: Agent-Assisted-Purchase-Research-Index collects everything about agent-driven purchase research — skills, templates, plugins and demos. Agentic-Purchase-Pipeline is listed there as a demo of that domain, and here as a reference workflow. Neither listing is redundant: one answers "what do I have for buying things?", the other answers "show me a workflow that was actually run, including where it broke."