--- name: production-quality description: Plan, produce and verify every new or materially revised Follow.fyi story as a complete reading and sharing experience before publication, including automated runs. --- # Production quality This is the release checklist for every article, including automated runs. Full stories default to an **illustrated narrative**: compelling documented storytelling, a substantial opening image, and purposeful images distributed through the developments they explain. Visual production is part of the assignment from commissioning through review, not an optional decoration after writing. Length, source count, image count and animation do not establish quality. Apply the reporting workflow's [Astra authoring requirement](../follow-story/SKILL.md#authoring-requirement) before drafting or revising any story text. It defines the authoring configuration, permitted supporting work, publication declaration and verification limits. ## Treatment standard Use [Your Attention Is Power. Who’s Using It?](https://follow.fyi/stories/your-attention-is-power-whos-using-it) and [SAP’s AI Test: Will the Moat Survive?](https://follow.fyi/stories/saps-ai-test-will-the-moat-survive), visual edition 3, as the owner’s reference treatments. Before planning a new full story, inspect an actual rendered reference or its saved rendered proof. Compare opening scale, image/prose rhythm, the explanatory work of images, and phone readability. These are craft benchmarks, not layouts, colors or image totals to copy. Record the reference inspected and which qualities the new story must deliver; a text-only extraction cannot establish a visual comparison. An illustrated narrative is complete when: - A substantive opening image establishes the story’s specific tension or question, rather than merely identifying the topic. - A coherent sequence of substantial illustrations, rights-cleared documentary images or authored explanatory graphics carries distinct developments across the story. Give images real reading space at their narrative anchors; a hero followed by an uninterrupted essay is incomplete. - Prose and images develop an opening, explanation, challenge and bounded ending together. Visuals expose a mechanism, comparison, consequence or change in understanding; repeated generic imagery does not fulfil the treatment. - Every image has a defined reader question, narrative anchor, evidence boundary, source IDs where applicable, useful alt text and visible caption/credit. Conceptual artwork is visibly identified as illustration, with retained prompts and provenance; it cannot impersonate an observed scene or establish a fact. - Charts, tables, timelines, source panels, icons and the social sharing card support the experience. They do not by themselves fulfil the opening image and distributed in-story image treatment. A necessary explanatory diagram may be part of that treatment; a screenshot of prose or a table is not a substitute. A genuine short evidence answer, narrow correction or explicitly requested format may use a proportional treatment. Record that scope and its reader-serving rationale when commissioning, or preserve a later explicit user change. Existing authorization remains valid; these rules do not introduce a new approval request. A short draft, deadline, cost, missing image tool or failed upload does not turn a full article into a brief. When a commissioned treatment cannot be completed, retain the draft and hold release with the concrete repair needed. A material revision preserves the published visual quality unless the verified account or explicit user direction warrants changing it. ## Work 1. **Plan the story before prose.** Read the commission, evidence verdict and strongest challenge. Start a visual beat plan before drafting the complete article: opening tension and qualified answer, consequential developments, explanatory or historical turns, strongest countercase, and bounded ending. For each beat, name the verified material, what changes in the reader’s understanding, prose’s job, image’s job, narrative anchor, and any visual evidence boundary. Choose the treatment under the standard above and check a rendered reference. Use `narrative-writing` for the story arc and `visual-journalism` for images, charts and mechanisms. Completion: the plan specifies the opening image and distributed visual sequence, or the genuine scope exception, and every beat advances the reader’s question rather than filling a template. 2. **Check delivery before promising it.** Inspect the current tools and publication schema for research, image generation, durable asset storage, supported visual types, preview, review and publication. An image generator returning a local file is only the first stage of delivery. Use `upload_story_image` for a newsroom or personal follow asset: prepare a supported raster file within the current tool limits, supply its base64 bytes with the newsroom request ID or personal follow ID as requestId and one UUID assetId, and retain the receipt containing its path, dimensions and checksum. Reuse the assetId for an uncertain retry; changed bytes need a new assetId. Use the returned dimensions and src in the image visual. A newsroom upload remains editor-only until a public edition references it. Personal follow uploads and their published personal editions remain private to their owner. Choose PNG, JPEG or WebP for a hero, within the current social renderer pixel limit; an in-story SVG diagram cannot serve as the sharing hero. Owner-authorized Site deployment is a separate route for supported assets that the upload tool cannot carry. If a required capability is unavailable, preserve the draft and report the missing capability; an automated run must hold the promised treatment rather than quietly publish a reduced one. A stale connector schema is a delivery failure: refresh discovery or use the documented authenticated publishing surface; report unresolved mismatch. 3. **Produce prose and images together.** Write the story in authored order with its own pace and section lengths. Complete the planned original or rights-cleared image assets as the story develops. Let image-led, prose-led and evidence-led passages do different work; remove redundant explanation rather than extending long blocks with decorative images. Revise the beat plan when the evidence changes. Keep source IDs on all factual sections, visuals and change notes; the reader encounters source ledgers, detailed methods and verification records in **Sources & evidence**. Keep attribution, uncertainty, definitions and caveats that change meaning in the prose or visible caption. A separate evidence tab improves reading flow; it never changes what the claim is allowed to mean. 4. **Finish the visual assets.** Inspect every delivered image at its intended crop and at phone width, in the actual reader. Confirm substantial scale, legible relationships and labels, coherent visual language, and meaningful distribution through the story. Documentary media needs provenance and rights. Generated conceptual art must be labelled as illustration and must not impersonate evidence of a real scene. Choose charts for measured comparisons and bounded models for mechanisms. Label assumptions and simulated values; use the current supported renderer, not arbitrary executable payloads. Check labels, alt text, keyboard use, textual/data fallback and reduced-motion behavior. Optimize assets and loading within the current performance contract; reduce file weight or unnecessary complexity without silently dropping the commissioned visual sequence. Completion: the same finding and limitation remain available without animation, hover or color recognition, and every referenced asset returns nonempty content with its expected media type in the authorised preview. Check public access again after release; private and held drafts must remain private. 5. **Prepare the share.** Use a truthful, specific headline and a concise deck that communicate the same account. For a public article, confirm its canonical public URL, Open Graph and Twitter title, description and image. For an illustrated story, the actual story-specific sharing image must incorporate its hero artwork in a truthful, legible crop, with provenance retained. The social card is a separate deliverable from the article’s opening image and visual sequence. Inspect the actual share image and its crop, including headline legibility when text is part of the image. After public publication, the image must load for an unauthenticated crawler. For a personal or private story, record sharing as intentionally private: verify owner-authorized reading and image access, and denial to anonymous readers and other owners. Do not create a public route or image to satisfy this checklist. A local pathname, text-only substitute or promised future upload does not complete an illustrated story’s share treatment. 6. **Review the exact edition.** Send the complete artifact, beat plan, rendered benchmark comparison, evidence and visual provenance to the independent `publication-review` stage. Bind findings to the edition's revision or content hash and final asset checksums. The reviewer checks the actual title, deck, prose, figures, captions, image assets and sharing metadata, and explicitly assesses the treatment standard above. Check that the opening creates a specific question, each development changes understanding, the countercase receives real weight, and the ending answers within the evidence. Inspect the images in order and the complete narrative flow; neither a visual count nor attractive isolated screenshots proves a finished story. Record specific repairs and remaining limits; any changed claim, number, label, caption or asset invalidates its affected review. Reader experience review must include a phone and desktop, the Story and Sources & evidence tabs, and meaningful interactive controls. A missing preview, independent review, required asset or distributed image treatment leaves a full story on hold, with the concrete remedy recorded. 7. **Release and read back.** Publish through the authorised revision-safe workflow. Compare the accepted edition with the reviewed artifact, then inspect the route, image responses and sharing metadata within the commissioned audience. Public stories require anonymous readback; personal/private stories require authorized readback and proof that unauthorized access is denied. Confirm that the evidence tab maps the story's claims and visuals to the intended sources. Preserve older published editions and identify material changes. Completion requires a publication receipt and a matching rendered edition in its authorized audience; a saved draft, upload or Site deployment alone does not establish publication. ## Record and outcomes Keep a `productionRecord` with the commissioned scope and treatment, visual beat plan, inspected benchmark and comparison, image purposes and placements, any genuine scope exception, capability gaps, asset provenance and URLs, exact review hash, mobile/desktop and accessibility checks, share check, and release/readback receipt. Include a substantive `method-production` section in each new or materially revised edition, recording the treatment, actual checks, assets, independent review result and any limits. Give it valid source IDs for the material reviewed. Populate the current publication schema’s structured `production` contract and review evidence, alongside the detailed productionRecord retained with the research. Use the current tools and repository production checker for its exact fields and canonical digest; never guess a hash recipe or copy a previous edition’s review. The method section belongs in Sources & evidence. Do not place the canonical digest inside that section or another hashed field; it belongs only in `production.review.editionSha256`, so it does not change its own input. The publication contract is `production.version: 1`, the `authoring` declaration defined in the reporting workflow, a treatment, benchmark account, `storyPlan` (opening, development, ending), a `visualPlan` entry for every delivered visual, and `review`. Use `prepare_publication_review` on the complete candidate to obtain the normalized digest and structural failures. After independent review, record its actual reviewer, timestamp, result, specific narrative/visual/desktop/mobile/source/keyboard/sharing checks and unresolved findings under `production.review`; use the returned digest. Re-run the preflight after attaching the review. The tool cannot inspect assets or perform that review for you. Any change outside the review record requires a new digest and renewed review. The repository checker is `scripts/check-production.mjs`; follow its current usage. The runtime rejects an illustrated edition with fewer than three distinct images, without exactly one supported raster hero, without body images at distinct narrative anchors, or without image distribution across both halves of a longer narrative. Every planned visual must exist at a narrative anchor. These checks catch omissions; the reviewer still judges scale, purpose, quality and the full visual sequence. The `brief` exception is limited to three narrative sections and 600 words (including title, deck and summary), requires a scope rationale, and must match a newsroom brief commission. The current runtime cannot change an existing newsroom commission from narrative/explainer to brief through a revision request. Preserve a later explicit user direction in the record and report that concrete capability gap; do not claim the conversion is supported, relabel the existing commission, or bypass its gate. A substantial correction or revision retains the full story treatment. Do not mislabel a long text draft as a brief to bypass production requirements. State which structural checks the live runtime enforces and which judgments were performed by the reviewer. A minimum image/anchor check only catches obvious omissions; it is a floor, not a target or certification of pacing, relevance or craft. Asset presence, a schema pass or a self-written pass flag cannot certify narrative or visual quality. Read compatibility for older editions does not excuse missing production evidence on a new publication. Use **publish** only when the reviewed treatment is complete. Use **revise** for repairable findings and **hold** when the reporting, review or delivery cannot yet support it. A no-change monitoring check preserves the published edition and its evidence cutoff. Automatic deadlines change how many stories can be finished, not the standard for any story that is released.