The problem
Every time Sharp Method closed a new engagement, the same manual work followed. A directory had to be created and structured by hand. A GitHub repo needed to be initialized. A CLAUDE.md context file had to be drafted with client-specific brand guidelines, audience notes, and open questions. Research, evidence, call, and proposal folders all had to be set up before an agent could do anything useful. None of this was hard — it was repetitive, error-prone, and unbillable.
The pipeline itself ran in a sensible order — lead, audit, call, recap, proposal, paid discovery, SOW, build — but that order was tribal knowledge, not a system. The gate between a signed SOW and the actual start of build work was porous: work sometimes started before payment cleared, kickoff was ad hoc, and context lived in different places rather than one canonical location.
The deeper cost was compounding. Each new engagement rebuilt the scaffolding from scratch. Any gap in the setup — a missing brand guideline, a folder in the wrong place, stale context from a previous run — degraded agent output downstream. When output drifted, the instinct was to re-prompt. The real fix was always the context.
The approach
The first design decision was structural: one repository per client, always. Everything the agent produces and everything it reads lives there. A well-maintained repo with a current CLAUDE.md produces consistent output; a repo with stale or incomplete context produces drift — which looks like a model problem until you realize it is a context problem.
The second decision was to gate the pipeline on a real business event. The build phase does not begin until the SOW is signed and the 50% upfront payment has cleared. This filters out non-serious engagements and gives the team a clean starting signal. Paid discovery — a fixed-fee, time-boxed period where the deliverable is a full SOW with fixed prices — serves the same function one stage earlier: it replaces unpaid scoping with an engagement that has a price, a hard end date, and full credit toward the build if the client proceeds. If they do not, they walk away with a build quote, migration plan, and site audit.
The third decision was what not to automate. The pipeline automates scaffold creation, context gathering, audit output, call transcription, and proposal drafting. It does not automate the final judgment calls: which recommendation to lead with, whether a claim is backed by enough evidence, whether the demo is ready to send. Those passes stay human. The value in this market is taste, and taste cannot be delegated to a pipeline.
The build
The pipeline runs as eight stages, labeled 00 through 07. Stage 00 is workspace initialization: git init, directory skeleton, and a populated CLAUDE.md with the client's company story, trigger, brand guidelines, audience priority, revenue model, and a glossary of trade terms. The agent reads this file at the start of every session; correcting it the moment outputs drift is the most critical maintenance task.
Stages 01 through 03 cover pre-call preparation: automated site audit (UI/UX, SEO, performance, competitive, brand), call brief generation, and post-call processing. The call brief is an HTML reference document with a prioritized question list, company snapshot, and industry glossary drawn from audit findings. Call recordings are transcribed locally using faster-whisper, then passed to an agent that updates the call sheet, flags unanswered questions, and drafts the recap email. Stages 04 and 05 cover proposal construction and the paid discovery period. Stage 06 is the SOW-and-payment gate. Stage 07 is the build phase, where agent goals with explicit exit criteria drive development in small increments with human review at each step.
The components that make stages mechanical rather than manual:
- A
_templates/directory with reusable call-brief, proposal, recap-email, and SOW templates - A
_prompts/directory with the audit prompt and variants, all versioned - A
_skills/library of custom agent skills adapted to the pipeline's folder structure - A bash transcription one-liner that converts any call recording to a timestamped transcript without leaving the terminal
Outcomes
The first full pipeline run — from lead arrival through a live proposal with interactive add-on pricing — was completed on an actual client engagement. The homepage demo was built in approximately two hours across fifteen to twenty model iterations. Artifacts delivered: a structured repo, populated CLAUDE.md, full audit across five disciplines, HTML call brief with prioritized questions and industry glossary, call transcript, updated call sheet, and a hosted proposal page with add-on toggles that live-update the project total.
The pipeline makes explicit what was previously implicit: where each stage ends, what artifact gates the next, and where money changes hands. The 50% payment gate is structural. Paid discovery is the default. When agent output drifts (wrong colors, off-brand copy), the fix is an update to CLAUDE.md, not a new prompt.
Both core deliverables — the site audit and the live demo — give the prospect immediate value before any money changes hands: pipeline outputs, not one-off heroics.
What's next
The current build encodes a review loop — pipeline drafts, first internal pass, partner pass, feedback, regenerate — before anything reaches a client. The next development is a lightweight dashboard showing each active client's current stage, the gate condition that needs to clear, and any context items flagged as stale. Each new engagement deposits reusable assets — call briefs, proposal structures, audit outputs — that make the next run faster.