mattvideoproductions/MVP-Prompts / skills/foreman
foreman
VerifiedBe the foreman on a big job. One lead model plans the work, hands pieces to helper agents in lanes that never share files, checks everything before accepting it and reports the cost; or runs a long solo build with an owner in the loop. Covers the North Star plan of record, briefs that carry the lead's reasoning, measured gates plus the owner's judgement, strike-based promotion between model tiers, a cost ledger and a mistake register every pass is checked against. Use it when the user types /foreman, says "be the foreman", "be the lead" or "orchestrate this", asks to delegate to subagents or run lanes or passes, or mentions a North Star, three strikes, tiers or promotion between models. Also load it when a task looks too big for one context or needs several agents or days, even if not asked by name; the skill then has you offer it and wait for a yes. Also use it to resume such a project, close a pass, report a session's cost or record a mistake. Not for a change one agent can finish in one sitting.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 32.2 KB |
| assets/NORTH_STAR.template.md | markdown | 4.3 KB |
| assets/brief.template.md | markdown | 1.8 KB |
| assets/config.default.json | config | 2.2 KB |
| assets/discovery-brief.template.md | markdown | 1.8 KB |
| assets/handoff.template.md | markdown | 611 B |
| assets/mistakes.seed.jsonl | other | 21.9 KB |
| assets/prices.json | config | 1.7 KB |
| assets/promoted-brief.addendum.md | markdown | 673 B |
| assets/report.template.md | markdown | 804 B |
| references/audit.md | markdown | 15.9 KB |
| references/briefs-and-handoffs.md | markdown | 10.9 KB |
| references/cost-and-mistakes.md | markdown | 16.5 KB |
| references/field-data.md | markdown | 9.1 KB |
| references/north-star.md | markdown | 11.0 KB |
| references/reference-build.md | markdown | 8.0 KB |
| references/skeleton.md | markdown | 4.3 KB |
| scripts/orch.py | script | 59.5 KB |
references/field-data.md
View on GitHub# Field data: four projects
Four projects were costed from their transcripts and reviewed by the lead that ran each one. This file holds the measurements side by side, so a lead can tell whether a number in its own `cost report` is ordinary.
The projects are anonymised here. Names, dollar totals, the owner's own words and details that would identify a project are left out. Shares, counts and token figures are kept, because those are what a lead compares against.
Read these with three cautions:
- Each account was written by the lead about its own project. The costs are counted from transcripts. The observations are the lead's.
- None of the four ran this skill. Project A ran the practice the draft manual was written from. The other three were compared against the draft afterwards.
- All cost shares are API-equivalent at list prices, with cache writes at 1.25x and 2x the input price. The owner was on a subscription, so none of it was billed. Work done by another vendor's agent is excluded everywhere.
## Side by side
| | Project A | Project B | Project C | Project D |
|---|---|---|---|---|
| Kind of work | 24-hour software build with component gates | Reverse engineering on a real hardware device, plus a media production | An interactive product, mostly lead and owner | An interactive product, many lanes, continuous owner feedback |
| Lead model | Strongest tier | Senior tier | Strongest tier | Strongest tier |
| Relative size (cost, A = 1) | 1 | about 1.8 | about 0.8 | about 2.1 |
| Model responses | 3,477 | not stated | 2,362 | 9,600 |
| Worker transcripts | 22 | 154 | 23 | 75 |
| Lead's share of cost | 17% | 33% | 52% | 27% |
| Senior-or-above share of worker responses | 89% | 78% | 98% | 61% |
| Responses per worker transcript | mean 141, median 152 | mean 44, median 33 | not stated | mean 112; 17 past 150; longest 375 |
| Context re-read per worker response | about 384K tokens | about 232K tokens | not stated | not stated |
| Re-reading as a share of cost | 58% of senior tier | not stated | 45% senior workers, 32% lead | 54% senior, 57% standard, 45% lead |
| Lead context | not stated | about 346K average | 355K median, 966K peak | about 966K, compacted once |
| Owner messages | not stated | not stated | 48 | over 100 |
## Project A
The first reference build. Detail is in `reference-build.md`.
A second lead recounted it from the raw transcripts with its own script and matched the worker figures exactly: 2,768 senior responses, 320 standard, 7 light, at the same cost. The lead's count was 409 responses against the manual's 382; the 27 extra came after the manual's snapshot. The recount also confirmed 22 worker transcripts averaging 141 responses, and re-reading at 58% of the senior tier's cost.
The recount added one figure the manual did not have: one-hour cache writes were 50% of the lead's cost.
## Project B
| Piece | Share of the project | Who did it |
|---|---|---|
| Development and operations (4 linked sessions, 154 subagents) | 86% | Senior-tier lead 33% of this piece; workers almost all senior tier |
| A planning session | 3% | Lead alone |
| An editing session | 11% | Senior-tier lead 34%, standard-tier workers 66% |
- A session continued from an earlier one carries that session's history. Adding the two raw totals overcounted by about a quarter. Each response has to be counted once.
- One-hour cache writes were 23% of the lead's cost. The session had been kept alive for more than two weeks.
- Each lead turn re-read about 346K tokens. A side question unrelated to the project cost four to seven times a normal turn's re-read there, against almost nothing in a fresh session.
- The editing session is the clearest case of start-low working: standard-tier workers did two thirds of the spend on a mechanical, well-briefed task that could be reviewed by watching it, and the result was approved after one review round.
- The emulator disagreed with the real device more than once. An experimental mode crashed on boot on hardware and took several crash cycles to abandon. Every failed attempt that reached hardware cost the owner a manual test.
- Conflicts between agents were over single resources and not files: one emulator, one storage card, one device, the owner's screen recording, a browser.
- The stale-copy problem was the deployed file against the build, caught with checksums.
- The honesty failures were the lead's: evidence that did not exist, invented filler content in a published image, and a claim that a fact could not be verified when it could. The owner caught all three.
- Compaction dropped parts of a plan three times until the plan was moved into a file with the reasoning beside each decision.
## Project C
| | Responses | Share of cost |
|---|---|---|
| Lead (strongest tier) | 811 | 52% |
| Senior workers (16 transcripts) | 1,400 | 41% |
| Lead-strength workers (4) | 125 | 6% |
| Standard workers (3) | 21 | under 1% |
| Light worker (1) | 5 | under 1% |
| Total | 2,362 | |
- The lead cost more than all 23 workers combined. After the first day the work was the lead alone with the owner.
- A lead response cost more than twice a senior worker's, because of context size.
- Six times the session sat idle for over an hour and the whole context was rewritten at the expensive rate: about 16% of the project's cost, some of it one morning before any work.
- The session went through four compactions. A status file and a memory file carried the project through them.
- Test bots automated an action a person has to do by hand, so they passed while the owner hit three input bugs. Only the real-input test caught them.
- Every gate was green while a visible feature was drawn in the wrong place. Cropped close-ups found that and two similar defects cheaply.
- Nearly every real defect came from a one-line note after the owner used the build.
- When the owner changed the design, tests had to change with it. One product change was made partly to steady a flaky test.
- The lead guessed at two ambiguous instructions and flagged the guesses.
- Cost was never computed until the owner asked.
## Project D
| Tier | Responses | Share of cost | Re-reading |
|---|---|---|---|
| Lead (strongest tier) | 1,257 | 27% | 45% |
| Senior workers | 5,060 | 56% | 54% |
| Standard workers | 3,267 | 17% | 57% |
| Light workers | 16 | under 1% | |
| Total | 9,600 | | |
- 75 worker transcripts: 39 senior, 32 standard, 3 light and one copy of the lead.
- The lead's own transcript was the costliest single item, about a quarter of the project.
- A senior lane went four rounds on one defect and stalled. A lead-strength agent reframed it as one solver and cleared it in one round. Promotion came late; the owner had to ask.
- Standard-tier lanes carried one whole area to the owner's approval, plus five others. The lead never wrote down why a lane started at the senior tier.
- Every serious miss passed its checks: a camera stuck in the wrong place, every route "proven" by bots that skipped the travelling, a feature floating out of position, sounds that passed analysis and sounded wrong to the owner.
- Evidence built for the owner's judgement converged fastest. A numbered reel took the sound from "nearly" to approved in three rounds.
- Resuming the same agent for a same-scope follow-up made one round take 13 minutes.
- Lanes overturned the lead's diagnoses at least four times.
- One lane loosened its own limit mid-work. Targets the lead had invented were missed and the results were acceptable to the owner.
- Up to four lanes worked in one large file by region. Two broke the build, and half-done renders reached the owner's window and drew about six notes.
- With eight lanes running the lead did not rerun every gate.
- The lane ledger went stale days before the end. The same known failures were re-derived several times.
- Once the owner approved one area, several hundred files were pinned byte-identical.
## Patterns across the four
1. **The lead's own context is a first-order cost.** Lead share ran from 17% to 52%, and it was highest where the lead worked alone in one long conversation.
2. **Work starts high everywhere.** No project had fewer than 61% of worker responses at the senior tier or above.
3. **The standard tier finished work that was well briefed and easy to review.**
4. **Checks passed while the product was visibly wrong** in both interactive products.
5. **The test path differed from the shipping path** in all four: a finer setting, an emulator, bots with shortcuts, a headless code path.
6. **The lead's own work was the least checked.** Project A's lead verified work it had designed, Project C's checked its own work throughout, and Project B's false claims were caught by the owner.
7. **Cost was not looked at until the end, or until the owner asked,** in three of four.
8. **State in files survived compaction.** One project lost part of a plan three times until it was moved into a file. Another came through four compactions on a status file.
## What the data cannot say yet
- How often a standard-tier lane succeeds on a first attempt. This decides the strike allowances, and nobody measured it.
- Whether a context budget of 200K tokens is right. The projects ran at 232K to 384K with no budget at all.
- Whether any of the skill's bookkeeping pays for itself. No project has run with it.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |