Social Publisher - AI publishing, opportunity intelligence, and creative production

Design the controls around human–AI collaboration.

I translated the AI workflow into product interactions: editable campaign strategy, inspectable candidate context, intent-adaptive reply drafts, content and media review, and separate publishing outcomes for each platform. Users can guide the strategy and review consequential outputs at the appropriate step.

I built request-level tracking for each generation stage, showing model usage, token consumption, completion status, and estimated cost.

I connected these records to budget reservations and platform diagnostics to identify which calls were expensive, incomplete, or failing.

Analyze the publisher once and reuse memory across content workflows.

I pre-analyzed user-provided profiles, portfolio material, project information, experience, and preferences, then stored the resulting context in persistent memory. Topic selection, post writing, creative direction, and replies reuse that context to represent the publisher consistently.

User corrections are saved as reusable preferences and rules. This keeps the publisher’s background and communication style available across jobs while reducing repeated analysis and unnecessary model input.

Connect strategy to discovery, qualification, and response.

The saved campaign guides automated discovery on Reddit or across the web. Candidates are assessed against audience fit, intent, problem fit, the campaign objective, freshness, and reply appropriateness. Results include Why engage, Watch, and an editable draft.

I made reply behavior adapt to intent: an explicit hiring or paid-project request can receive a relevant introduction; a solution-seeking discussion gets a helpful answer with an appropriate connection to the user’s work; general discussions stay expertise-led or value-first. Replies remain available for human review.

Separate source truth, AI judgment, and human authority.

I designed discovery → availability and freshness checks → qualification → intent-adaptive drafts → human review. Uploaded case studies are analyzed, reviewed, and saved as reusable project information. User corrections become independent, persistent rules.

A compact card is for scanning; reasoning uses the full canonical conversation, including external URLs and publication time. This distinction fixed drafts that asked for links already present or omitted a stated budget. Creative direction and factual supporting information are separate inputs.

Choose providers and architecture by task fit and failure behavior.

Provider judgment was a practical workflow decision: which stages need semantic discovery or creative generation, and which are better served by deterministic software? I compared actual outputs, intent detection, completion behavior, integration constraints, and recorded cost rather than relying on a polished demo.

I chose search and generation approaches by task fit, output quality, completion behavior, integration requirements, and measured usage.

AI vendor evaluation: OpenAI and Meta APIs.

I compared OpenAI API and Meta API performance and cost in Social Publisher’s workflows. I assessed the usefulness of the outputs alongside API spending, using task fit and practical operating cost to guide provider choices.

I treated AI-provider selection and search architecture as separate decisions: comparing OpenAI and Meta addressed generation quality and cost; comparing AI web search with Reddit-native/local retrieval addressed discovery quality and token usage.

Trace the failing stage before changing the model or the code.

I separated creative quality, model completion, integration failures, and local accounting. A bad image, an incomplete caption, and an unpublished post require different investigations. The useful unit of diagnosis is the workflow step and its returned status, usage, and output—not a generic “AI failed” message.

I checked published copy and media on the destination platforms, connecting API diagnostics to the actual user-visible result.

Measure where tokens go; preserve quality and completed work.

One failed run reported 137,685 tokens already used plus 68,692 requested: a potential 206,377 against a 200,000 token-per-minute limit. This was a rate-budget failure, distinct from a model context-window limit or an unexpectedly high dollar charge.

I measured usage before changing the workflow. An initial 24,914-token run attributed 22,530 tokens to discovery and 2,384 to qualification and drafting. Recording calls by stage showed that discovery—not the short final replies—was the main token cost. Cached tokens were included within the total usage.

A freshness/resilience run returned four reviewable opportunities but consumed 146,944 tokens, including 136,850 for discovery. Breaking usage down by stage exposed repeated discovery as the source of the cost regression.

Choose search quality over lower token usage.

After campaign strategy, I explored a retrieval-first alternative: use Reddit-native/local search to obtain candidates, filter and deduplicate them, then pass a smaller result set and necessary source context to AI for qualification and drafting. The intent was to move structured discovery work out of expensive model-driven search.

I compared local/Devvit retrieval with AI web search by reviewing whether the candidates contained useful hiring and buyer intent.

The restored search baseline used 25,030 tokens. A subsequent intent-adaptive run used 24,609 tokens and produced 11 reviewable results. I reviewed candidate relevance and reply behavior throughout the iteration.

I retained the lower-cost experiment and its negative result. The lesson was to optimize the cost of useful, reviewable opportunities—not raw token count alone.

Keep the regressions and the source context visible.