The 825 resource library

Breadcrumbs

Practical workflows and guides. Find a process, understand its limits, and adapt it to your work.

All resources · Breadcrumbs · StaleToast

Reviewed source snapshot imported 2026-10-20. This is not a live feed. Source and contributions.

Structured repository context pack

Agents repeatedly rediscover architecture, conventions, and known traps.

Maintain a bounded index, architecture map, glossary, conventions, patterns, gotchas, area flows, and validation reports.

Evidence: public-synthetic-reproduction. Recorded: 2026-09-01.

Open workflow guide

Quick read

A new person or agent needs to orient to a repository without reading every file.

Write one index linking architecture, file map, conventions and known pitfalls.
Give each file a clear scope and keep detailed evidence behind links.
Run the context-budget guard and check every linked path.
Ask a fresh reader to identify the next safe task, then fix any missing route.

Check your result

The links and budgets pass, and a fresh reader can name the next action without guessing.

Limits

AI-maintained context can preserve errors unless evidence and review are required

Context-budget guard

Original source

Git-managed scoped agent memory

Agent memory becomes opaque, over-broad, or difficult to correct.

Store atomic facts in explicit persona, shared, and workspace scopes with on-demand recall and Git review.

Evidence: public-synthetic-reproduction. Recorded: 2026-09-01.

Open workflow guide

Quick read

Facts need to survive sessions while keeping scope, corrections and sources visible.

Separate facts from tasks and decisions.
Give each fact its source, scope and correction history.
Retrieve only the scope needed for the current question.
Test expected recall and forbidden cross-scope recall with synthetic facts.

Check your result

Current facts are retrieved, superseded facts are identified, and forbidden scopes stay out of the response.

Limits

Git history alone does not establish truth or authorization

Scoped-context tests · Memory desk

Original source

Structured SDLC rules and skills

General agent instructions are too broad to enforce task-specific quality gates.

Express bounded development stages as task-specific rules or skills with guardrails and specialist roles.

Evidence: public-synthetic-reproduction. Recorded: 2026-09-01.

Open workflow guide

Quick read

A repeatable development task needs explicit entry, review and completion gates.

Define the task, allowed changes and done condition.
Name the author, reviewer and release owner; give each a bounded role.
Work one stage at a time and attach test evidence before advancing.
Stop at an unresolved permission or evidence gap and record the next action.

Check your result

Every completed stage has its evidence, and unfinished gates are visible rather than silently skipped.

Limits

more roles can add coordination cost without measured benefit

Rules spine

Original source

Small-scope implementation and handoff

Long sessions blur completion state and accumulate unverified assumptions.

Keep scope small, define done before editing, run tests, then compact or hand off before starting another task.

Evidence: public-synthetic-reproduction. Recorded: 2026-09-01.

Open workflow guide

Quick read

A change or handoff needs a clear completion record before another task starts.

Write the objective, scope and completion check before editing.
Make the smallest change that satisfies that objective.
Run the relevant checks and retain failures as well as passes.
Hand off the source revision, result, unresolved items and next owner.

Check your result

Another reader can verify the result and continue from the handoff without inferring that a publish or deployment happened.

Limits

small slices still require a coordinator to preserve program-level coherence

Session handoff template

Original source

Design-to-code cross-model bridge

Design reasoning and implementation execution lose context across tools.

Use a structured brief to transfer approved design intent into a bounded implementation session and return code evidence through a gate.

Evidence: public-synthetic-reproduction. Recorded: 2026-09-01.

Open workflow guide

Quick read

An approved design needs to pass between reasoning and implementation tools.

Prepare a short brief with objective, constraints, source revision and acceptance checks.
Have the receiving owner acknowledge the scope and missing evidence.
Implement within that scope and return an exact diff and test results.
Review the result against the original brief before changing an adoption or release state.

Check your result

The returned implementation matches the approved brief, and any departure or unverified result is explicit.

Limits

a handoff does not prove the receiving model understood or completed the task

Workflow worksheets

Original source

Pause work for approval, then resume the approved revision

A long-running job needs a human decision before its next consequential step.

Prepare the result, pause at a review checkpoint, and resume only after an explicit decision about the same revision.

Evidence: pinned-primary-documentation-screen. Recorded: 2026-10-13.

Open workflow guide

Quick read

A draft or job output is ready, but publication or another consequential next step still needs an authorized human decision.

Prepare the draft and identify its exact version. Keep the consequential action paused.
Show the result and the proposed next action to the decision owner. Record a deadline.
Collect approve, reject or request-changes against that version. In a running system, a wait token can carry the response; a manual queue can record the same decision.
Before resuming, check that the owner, version and deadline still match. Changed work needs a fresh decision.
Resume the approved step once, or leave rejected and expired work paused. Record the outcome and a correction path.

Check your result

In a synthetic rehearsal, only a current approval for the reviewed version permits the next step. Rejection, expiry, wrong owner, changed version and repeated responses produce no extra action. This rehearsal is proposed, not completed.

Limits

The source describes a runtime mechanism, not proof of reviewer identity or decision quality.
The Trigger.dev implementation requires its SDK and service; the manual variant below does not reproduce that runtime.
Revision binding, reviewer authority and effect deduplication below are local safeguards proposed by this guide, not guarantees established by the upstream token API.
No upstream runtime, installation, local reproduction or consumer adoption was performed.

Pinned upstream review example · Stage gates guide · Workflow evidence contract

Original source